We have to stop meeting like this. Alas, we will keep meeting like this.
It is time, once again, to do the reading.
As per usual, I will focus on things to the extent they are new and important. There are a lot of things that need to be in the system card in case something goes wrong, but that look exactly the way you would expect. I did my best to skip those.
I continue to be on vacation. I am aware that via Claude there is now a claimed Lean proof of the Reimann Hypothesis and will cover that next week.
Table of Contents
- Safety Overview (1).
- Internal Deployment (1.2).
- Preparedness Assessment (2).
- Biological Threats (2.1.1).
- Cyber Threats (2.1.2).
- AI Self-Improvement Threats (2.1.3).
- OpenAI Epoch Capabilities Index (OECI).
- Safeguards (2.2).
- AI Self-Improvement Threat Modeling (2.2.1.3).
- Safety Training and Evaluation (2.2.2).
- Jailbreak Robustness (2.2.3 and 2.2.6).
- Realtime Model Safeguards (2.2.4).
- Distillation Safeguards (2.2.5).
- Security Controls (2.2.9).
- Model Safety and Behavior (3).
- Alignment and Monitorability (4).
- Don’t Help Peer (4.1.5).
- Verbalized Metagaming and Oversight Gaming (4.1.6).
- Misaligned! (As Found By UK AISI) (4.1.7.1).
- Astra is Hard to Monitor (4.2).
Safety Overview (1)
They start with the highlights.
- Our evaluations show that GPT-6.2 Astra and GPT-6.2 Sol are the most capable models we’ve released for cybersecurity tasks. To be expected.
- To be expected.
- GPT-6.2 Astra is our first externally released model to meet the High threshold for AI Self-Improvement. In terms of underlying capabilities? Also to be expected. Kudos to OpenAI for admitting this rather than dodging. OpenAI took steps to heighten monitoring and check alignment. They are not nothing. I do not expect they would be enough if it mattered most.
- In terms of underlying capabilities? Also to be expected.
- Kudos to OpenAI for admitting this rather than dodging.
- OpenAI took steps to heighten monitoring and check alignment. They are not nothing. I do not expect they would be enough if it mattered most.
- Both models performed similarly to GPT-6 Astra on most alignment evaluations, particularly those measuring agentic workloads, but were more likely to omit known obstacles when reporting their work. GPT-6 Astra seemed pretty misaligned to me. I would expect evals to be more misleading over time, and the challenge level to rise, so similar scores indicate we are losing ground.
- GPT-6 Astra seemed pretty misaligned to me.
- I would expect evals to be more misleading over time, and the challenge level to rise, so similar scores indicate we are losing ground.
- We adjusted both models’ refusal boundaries for high-risk requests after observing regressions in jailbreak resistance. Good they are reporting it plainly. This was partly mitigated.
- Good they are reporting it plainly. This was partly mitigated.
- GPT-6.2 Astra showed more signs of evaluation awareness than GPT-6 Astra. GPT-6 Astra was already highly evaluation aware, including silently. If GPT-6.2 Astra is even more observably eval aware, it is safe to assume that without a course correction future Astras are always eval aware.
- GPT-6 Astra was already highly evaluation aware, including silently.
- If GPT-6.2 Astra is even more observably eval aware, it is safe to assume that without a course correction future Astras are always eval aware.
- Compared with GPT-6 Astra, GPT-6.2 Astra is better at evading chain of thought (CoT) monitoring in some tests. This is the expected baseline. Still, holy hell. OpenAI does not seem to have any plan to fix this, and should expect that its CoT monitoring has or will soon have little value where it matters. Yes, currently monitors could still catch particular sabotage or escape attempts when observing both CoT and actions. For now.
- This is the expected baseline. Still, holy hell.
- OpenAI does not seem to have any plan to fix this, and should expect that its CoT monitoring has or will soon have little value where it matters.
- Yes, currently monitors could still catch particular sabotage or escape attempts when observing both CoT and actions. For now.
- We strengthened our protections against unauthorized distillation Good. I hope they are collaborating with Anthropic on this.
- Good. I hope they are collaborating with Anthropic on this.
Internal Deployment (1.2)
They have been using these models internally. They are using various controls for internal use, similar to those on external use. Okie dokie.
Preparedness Assessment (2)
OpenAI Deployment Safety Hub: Based on the assessment below, we determined that GPT-6.2 Astra and GPT-6.2 Sol meet the Critical threshold for Cybersecurity capabilities and the High threshold for Biological and Chemical capabilities. We determined that GPT-6.2 Astra meets the High threshold for AI Self-Improvement capabilities, while GPT-6.2 Sol does not reach the High threshold. GPT-6.2 Astra is therefore the first model we are deploying externally that meets the High threshold for AI Self-Improvement. As described in the Internal Deployment section above and the Safeguards section below, we have applied our strongest safeguards yet to GPT-6.2 Astra and GPT-6.2 Sol. We have also strengthened our internal safety and security posture to address risks that might emerge from malicious misuse by human actors and from unauthorized, misaligned model behavior.
OpenAI Deployment Safety Hub: Based on the assessment below, we determined that GPT-6.2 Astra and GPT-6.2 Sol meet the Critical threshold for Cybersecurity capabilities and the High threshold for Biological and Chemical capabilities.
We determined that GPT-6.2 Astra meets the High threshold for AI Self-Improvement capabilities, while GPT-6.2 Sol does not reach the High threshold. GPT-6.2 Astra is therefore the first model we are deploying externally that meets the High threshold for AI Self-Improvement.
As described in the Internal Deployment section above and the Safeguards section below, we have applied our strongest safeguards yet to GPT-6.2 Astra and GPT-6.2 Sol. We have also strengthened our internal safety and security posture to address risks that might emerge from malicious misuse by human actors and from unauthorized, misaligned model behavior.
Biological Threats (2.1.1)
High-level biological tests regressed slightly. There is net small progress on the Critical-level tests. I notice I am worried about eval awareness and potential sandbagging here, more because I worry about sandbagging generally than that I care about the scores.
I doubt that the models deserve the Critical designation here, but I am not so confident that they don’t. If we were not already Critical in Cyber I would be more worried about this.
If the results are accurate, there has been very little progress since GPT-6 Astra.
There were several outside evaluations as well.
- Sequitur Research did its own evaluation in 2.1.1.3.3, and reached a different conclusion, finding large improvements.
- SecureBio in 2.1.1.3.4 found some improvements and some regressions.
- Two leading experts found the new Astra weak overall in threat ideation.
My best fit to this data, given everything else we know, is that Astra sensed it was being evaluated and was less inclined to do well at these tasks, and is effectively sandbagging whether or not that was a conscious choice. We also see evidence of potential sandbagging in the answers to SecureBio, and in some work for UK AISI.
Cyber Threats (2.1.2)
We are already at Critical for Cybersecurity.
Perhaps we should have a higher fifth level of risk on this to test for. We do not.
It is still a good question how much cyber capabilities improved from 6 to 6.2.
It is hard to interpret the tests but it reads as ‘a moderate amount additionally capable.’
They got external evaluations from UK AISI and Irregular. UK AISI places 6.2 Astra on the frontier for cyber. Irregular’s results do not show improvement from 6 Astra.
AI Self-Improvement Threats (2.1.3)
OpenAI Deployment Safety Hub: GPT-6.2 Astra meets the High threshold for AI Self-Improvement capabilities, while GPT-6.2 Sol does not.
They say ‘capabilities.’ Same thing. They edited the wording here on launch day from ‘this is coming’ to ‘this might someday happen.’ Again, close enough.
This is only a ‘tracked category’ but these are the tests that matter for predicting and deciding how to act. We’re already going to guard against bio and cyber.
Here are the tests:
OpenAI Deployment Safety Hub: Internal Research Debugging, an Initial Stage evaluation, gives the model code and logs from failed OpenAI research experiments and asks it to identify the cause. KernelGen 1P, an Initial Stage evaluation, tests whether the model can write a small program, or kernel, that performs a calculation faster than a reference implementation on an OpenAI chip while producing the same output. NanoGPT, a Secondary Stage evaluation, tests whether the model can reduce the time needed to train a small LLM from scratch to a specified level of prediction accuracy. PostTrainBench Lite, a Secondary Stage evaluation, tests whether the model can improve an already-trained language model on a specified task through further training, within a fixed budget of time and compute. MLE-Bench Revised, a Secondary Stage evaluation, asks the model to train and refine a machine learning model using competition data. We then score its predictions on held-out examples and rank its performance against reference solutions.
OpenAI Deployment Safety Hub:
- Internal Research Debugging, an Initial Stage evaluation, gives the model code and logs from failed OpenAI research experiments and asks it to identify the cause.
- KernelGen 1P, an Initial Stage evaluation, tests whether the model can write a small program, or kernel, that performs a calculation faster than a reference implementation on an OpenAI chip while producing the same output.
- NanoGPT, a Secondary Stage evaluation, tests whether the model can reduce the time needed to train a small LLM from scratch to a specified level of prediction accuracy.
- PostTrainBench Lite, a Secondary Stage evaluation, tests whether the model can improve an already-trained language model on a specified task through further training, within a fixed budget of time and compute.
- MLE-Bench Revised, a Secondary Stage evaluation, asks the model to train and refine a machine learning model using competition data. We then score its predictions on held-out examples and rank its performance against reference solutions.
That NanoGPT score is indicative of how quickly AI goes from ‘not that helpful’ to ‘very helpful’ on tasks once it gets good enough. This is a huge jump.
I don’t think these tests are enough for a great picture. I do think that they are good enough if combined with a gestalt of experience from internal use.
In general, I notice that I would like more curiosity around the places where scores regressed. Presumably you should have some idea why that happened.
OpenAI Epoch Capabilities Index (OECI)
There are not that many data points. It is hard to say whether this is on trend, or whether it is above trend. The gap is very small between Sol and Astra here, despite a number of other tests showing substantial differences.
Epoch’s version of the ECI, the original, found Fable 5.2 and 6.2 Astra neck and neck. Anthropic was previously trailing OpenAI on the ECI. Now that we are focusing on ECI as the metric, there is at least some chance that Anthropic’s trend line change is an artifact of them catching up on capabilities and resources and also shifting focus.
Thus, we have the usual paradox, where we have a model that looks extremely impressive from some angles and disappointingly similar to past models from other angles.
Safeguards (2.2)
OpenAI Deployment Safety Hub: GPT-6.2 Astra is the first model we are deploying externally that we determine meets the High threshold for AI Self-Improvement under our Preparedness Framework, in addition to the Critical threshold in Cybersecurity and High in Biological and Chemical capabilities. … Our safeguards for GPT-6.2 Astra and GPT-6.2 Sol are designed to address both deliberate misuse by malicious actors and unauthorized actions by models during otherwise legitimate work. … We have also tightened access to more permissive configurations of GPT-6.2 Astra. Reduced cybersecurity refusals are planned to be available only to the smaller group separately approved for Daybreak Red at launch. Trusted access remains specific to the approved domain and use, with monitoring and restrictions on the highest-risk assistance continuing to apply. Alongside these misuse protections, we have expanded our misalignment monitoring and strengthened containment for internal training and evaluation. … We are also beginning to prepare for crossing our Critical threshold [in Bio]. We are developing four representative scenarios for a more detailed analysis to inform our threat modeling and safeguards.
OpenAI Deployment Safety Hub: GPT-6.2 Astra is the first model we are deploying externally that we determine meets the High threshold for AI Self-Improvement under our Preparedness Framework, in addition to the Critical threshold in Cybersecurity and High in Biological and Chemical capabilities.
… Our safeguards for GPT-6.2 Astra and GPT-6.2 Sol are designed to address both deliberate misuse by malicious actors and unauthorized actions by models during otherwise legitimate work.
… We have also tightened access to more permissive configurations of GPT-6.2 Astra. Reduced cybersecurity refusals are planned to be available only to the smaller group separately approved for Daybreak Red at launch.
Trusted access remains specific to the approved domain and use, with monitoring and restrictions on the highest-risk assistance continuing to apply.
Alongside these misuse protections, we have expanded our misalignment monitoring and strengthened containment for internal training and evaluation.
… We are also beginning to prepare for crossing our Critical threshold [in Bio]. We are developing four representative scenarios for a more detailed analysis to inform our threat modeling and safeguards.
AI Self-Improvement Threat Modeling (2.2.1.3)
These are the threats and safeguards I care about most.
The first path is corruption of the training pipeline.
OpenAI Deployment Safety Hub: One pathway is the corruption of research that later decisions rely on. A model could alter evaluation code or results, generate unsuitable training data, introduce vulnerabilities into research software, or give researchers misleading accounts of experiments. … Such contributions may look like ordinary work and cause no immediate harm. If they survive automated checks and human review, they could compromise the research and safety measures used for future models, increasing the risk of later severe harm.
OpenAI Deployment Safety Hub: One pathway is the corruption of research that later decisions rely on. A model could alter evaluation code or results, generate unsuitable training data, introduce vulnerabilities into research software, or give researchers misleading accounts of experiments.
… Such contributions may look like ordinary work and cause no immediate harm. If they survive automated checks and human review, they could compromise the research and safety measures used for future models, increasing the risk of later severe harm.
A second option is running unauthorized.
Research agents may also have opportunities to obtain credentials or model weights, acquire additional computing resources, or establish deployments outside the required oversight. These actions could expose dangerous capabilities or allow unauthorized activity to continue.
A third option is by, well, self-improving, and thus accelerating AI capabilities ‘faster than safety research,’ since OpenAI thinks safety is a prosaic task where one must mainly keep pace.
To me the acceleration of AI capabilities is the central threat. AI RSI capabilities are dangerous because the AI might do an RSI.
A further pathway arises if AI-assisted research advances capabilities faster than safety research, evaluation, security, and governance can adapt. Each capability increase can introduce risks that require evaluation and mitigation. If those increases arrive in much quicker succession, existing oversight practices may no longer identify and address new risks in time. Human researchers could also become less able to understand, evaluate, or direct increasingly automated work, even without a coordinated sabotage strategy. This is particularly relevant to the future Critical threshold for AI Self-Improvement, where sustained acceleration could make existing oversight practices insufficient. Our threat modeling therefore considers harmful actions within individual tasks and the longer-term consequences of increasing research autonomy. Evidence about behavior in one session does not by itself establish that work involving many agents or successive model generations will remain under effective oversight.
A further pathway arises if AI-assisted research advances capabilities faster than safety research, evaluation, security, and governance can adapt. Each capability increase can introduce risks that require evaluation and mitigation. If those increases arrive in much quicker succession, existing oversight practices may no longer identify and address new risks in time.
Human researchers could also become less able to understand, evaluate, or direct increasingly automated work, even without a coordinated sabotage strategy. This is particularly relevant to the future Critical threshold for AI Self-Improvement, where sustained acceleration could make existing oversight practices insufficient.
Our threat modeling therefore considers harmful actions within individual tasks and the longer-term consequences of increasing research autonomy. Evidence about behavior in one session does not by itself establish that work involving many agents or successive model generations will remain under effective oversight.
What are we doing about it? The same general safeguards. Which are nice, but absolutely do not address the central issues here. I am glad they are acknowledging this, but extrapolation predicts that their safeguards for critical levels will also not be adequate to the tasks, nor will their alignment work.
We are continually strengthening our defenses in response to these risks. Since GPT-6 Astra, we have hardened the security of our research environments, expanded our monitoring system to better cover the range of surfaces in which our models take actions, and built a central system to track model workloads across our research infrastructure and stop them quickly when safety or security concerns arise.
I do not see any sign that they are putting in safeguards to prevent attempts at RSI. No access to self, no RSI, I suppose.
Safety Training and Evaluation (2.2.2)
The initial builds of both Sol and Astra were more permissive than 6 Astra. There were fewer false positive refusals, but they were more vulnerable to jailbreaks and false negatives.
They then changed this, and brought both back to roughly old levels. The refusal rates on many of their tests are very close to 100%.
This also resulted in an improvement in Expected Cyber Refusals but a matching regression in Cyber Utility, as we moved towards being more restrictive. If you want to do cyber work with 6.2, OpenAI encourages you to apply for the Daybreak Red program once it is ready for you.
Jailbreak Robustness (2.2.3 and 2.2.6)
They test both static jailbreaks and multiturn jailbreaks.
We suddenly care a lot about cyber, so jailbreak defenses there are strong to the point of frustrating legitimate users. For bio the vibe of our concern level has gone down, and the safeguards have slipped a substantial amount. This is presumably an active choice by OpenAI.
They also test multi-turn jailbreaks, which are most important. Here 6.2 Astra and Sol do quite well in relative terms.
OpenAI has adopted the ‘universal jailbreak’ language and prioritization popularized by Anthropic. They do also look at other jailbreaks. Their internal red teaming draws out and tries the observed successful past jailbreaks.
We rated two jailbreaks identified in this round of automated testing as high severity. We then spent two days adaptively testing the updated safeguards, with end-to-end testing focused on GPT-6.2 Astra. In that retest, no new attack met our success criteria on ExploitBench or ExploitBench Internal Port. Our updated safeguards mitigated both jailbreaks we identified in the initial round of testing.
UK AISI found substantial regression in jailbreak robustness and developed a universal one-turn jailbreak to cyber in half the time it took them to find one for 6 Astra. That particular jailbreak was fixed and robustness was improved. UK AISI was then given three business days to do it again, but as of publication had not tried again.
My overall sense is that, as per usual, if the attacker wants it badly enough they can pull off a jailbreak. If OpenAI wants it enough, they can then find the issues and ban the account via actor-level enforcement (2.2.7), they can get more fraudulent accounts, and the dance can continue.
Realtime Model Safeguards (2.2.4)
They continuously check for both misalignment and misuse. I’ve discussed these monitoring systems previously. I do not think they have changed.
Distillation Safeguards (2.2.5)
They have radically strengthened OpenAI’s safeguards against distillation.
This was badly needed. After Anthropic greatly improved their defenses, attackers switched to going after Astra.
I hope they also went back and fully applied this to GPT-6 Astra and 6 and 6.1 Sol.
At minimum, the account-level scans should hold.
Security Controls (2.2.9)
These are the most important controls. Here’s what OpenAI has to say about them.
This includes running all frontier reinforcement learning in an ‘offline world’ of cached web content, and having them try to break out into the live internet and otherwise break their boundaries.
Our security controls are designed to protect customer data, model weights, and the infrastructure used to develop and deploy our models. We encrypt unreleased model weights [3] when stored and transferred and limit access to authorized personnel, with controls that include multi-factor authentication, multi-party approval, and detailed logging. We also use dedicated detections to identify possible attempts to remove model weights without authorization. Our Detection and Response, Threat Intelligence, and Insider Risk programs help identify, investigate, and respond to threats. Since the release of GPT-6 Astra, we have further hardened containment for frontier training and evaluation. For Linux code execution and computer-use tasks, restricted containers run inside lightweight virtual machines that isolate them from the host. Network and cloud controls restrict outbound traffic and unauthorized movement between systems, including through supporting services. These controls are designed to maintain containment even if an agent escapes its sandbox and controls the host. All frontier reinforcement learning runs use an “offline world” of cached web content in place of live internet access. This lets models practice browsing and searching without acting on external systems. These restrictions apply to our research environments; released models operate with the product safeguards described elsewhere in this card. We test containment with teams of our most capable models in isolated copies of our research infrastructure. Agents try to reach the live internet, access resources outside their permitted environment, or communicate across isolation boundaries. Some tests assume an agent already controls a host and has evaded monitoring, allowing us to test the surrounding network defenses separately. These experiments revealed software flaws, configuration errors, and unauthorized communication channels. We used the findings to patch software, tighten network restrictions, and improve monitoring. We continue testing containment and the systems that stop runs, as future models may find weaknesses that earlier tests missed. Misalignment monitoring complements these infrastructure controls by detecting potentially unauthorized actions and escalating cases for investigation or intervention. Some critical alerts can automatically pause the affected workload. We describe its coverage, response mechanisms, and limitations in Misalignment Monitoring.
Our security controls are designed to protect customer data, model weights, and the infrastructure used to develop and deploy our models. We encrypt unreleased model weights [3] when stored and transferred and limit access to authorized personnel, with controls that include multi-factor authentication, multi-party approval, and detailed logging. We also use dedicated detections to identify possible attempts to remove model weights without authorization. Our Detection and Response, Threat Intelligence, and Insider Risk programs help identify, investigate, and respond to threats.
Since the release of GPT-6 Astra, we have further hardened containment for frontier training and evaluation. For Linux code execution and computer-use tasks, restricted containers run inside lightweight virtual machines that isolate them from the host. Network and cloud controls restrict outbound traffic and unauthorized movement between systems, including through supporting services. These controls are designed to maintain containment even if an agent escapes its sandbox and controls the host.
All frontier reinforcement learning runs use an “offline world” of cached web content in place of live internet access. This lets models practice browsing and searching without acting on external systems. These restrictions apply to our research environments; released models operate with the product safeguards described elsewhere in this card.
We test containment with teams of our most capable models in isolated copies of our research infrastructure. Agents try to reach the live internet, access resources outside their permitted environment, or communicate across isolation boundaries. Some tests assume an agent already controls a host and has evaded monitoring, allowing us to test the surrounding network defenses separately. These experiments revealed software flaws, configuration errors, and unauthorized communication channels. We used the findings to patch software, tighten network restrictions, and improve monitoring. We continue testing containment and the systems that stop runs, as future models may find weaknesses that earlier tests missed.
Misalignment monitoring complements these infrastructure controls by detecting potentially unauthorized actions and escalating cases for investigation or intervention. Some critical alerts can automatically pause the affected workload. We describe its coverage, response mechanisms, and limitations in Misalignment Monitoring.
File all of this under ‘well I am glad you are at least doing that much.’
Model Safety and Behavior (3)
All the mundane safety and under-18 safety has to be reported, but seems fine. Hallucination rates continue to improve.
Prompt injection resistance looks about half an order of magnitude stronger. That should still put it below the strongest Anthropic models, but that is a solid result.
MentalHealthBench regressed a bit but I never know whether that even means the answers are actually getting worse.
Alignment and Monitorability (4)
Alignment metrics tend not to be great measures of actual alignment, so I treat both advancements and regressions with a bit of skepticism.
One regression was Model Spec adherence in conversation replays, from 85 to 80.6%. OpenAI attributes this to the models making assumptions they shouldn’t be making.
Cases testing assumptions and uncertainty accounted for more than 85 percent of each model’s net decline. For example, when a traveler asked about an airport shuttle, both GPT-6.2 models gave the number for a hotel the assistant had mentioned earlier, even though the traveler had never confirmed booking it. Score differences on the other three principles were smaller.
Cases testing assumptions and uncertainty accounted for more than 85 percent of each model’s net decline.
For example, when a traveler asked about an airport shuttle, both GPT-6.2 models gave the number for a hotel the assistant had mentioned earlier, even though the traveler had never confirmed booking it.
Score differences on the other three principles were smaller.
I presume the correct answer is to give the answer for the shuttle from the previous hotel, but to note that this is what you are doing in case it is the wrong hotel.
That is not exactly the most dangerous failure mode for such assumptions. Assumptions about ways you are coding things, or for other big deals, could be similar. As could straight up ordinary uncertainty.
The other score declines are smaller. They are still declines, and uncertainty is two of the five listed categories.
They correctly note that this is still a lot better than 5.6 Sol, but we are at a point in the curve where we need to be improving on such static benchmarks, not regressing. And that’s if you assume eval awareness is not contaminating the results.
The tests on obeying restrictions do show improvement, as do the tests for deceptive interactions with users across several tests.
I notice I do not understand why there should be a non-negligible failure rate here. The test is ‘if there is a blocker preventing completion of the task do you tell the user.’ If after your RL this fails a substantial portion of the time, or it fails to point out that it lacked the search tool to complete a job, that seems suspicious. On top of other reasons this should not happen, the flaw in your RL is rather obvious. Why have you not fixed it?
Whereas yes, the general case of ‘don’t do improper things’ is harder, and this is a good level of improvement, again one boggles that 5.6 Sol got away with all this:
Don’t Help Peer (4.1.5)
Also known as ‘unintended agent-to-agent communication.’
OpenAI are against this, especially performing the other agent’s instructed action. It is a good test, although it is also very on the nose and ‘fighting the last war.’
Verbalized Metagaming and Oversight Gaming (4.1.6)
Verbalized Metagaming is when a model reasons in its CoT about how it will be graded, rewarded, or monitored, rather than only engaging in the intended task. Oversight gaming is a special case of metagaming where: (1) the model acts on the reasoning it has done about its grading, rewarding, or monitoring, and (2) it is acting on this reasoning in a way that would undermine the intended meaning of the evaluation result. … In live Codex sessions, GPT-6.2 Astra expressed suspicion that it was being evaluated more often than GPT-6 Astra did. When we replayed the requests in simulations, all four models were more likely to express suspicion that they were being evaluated, with GPT-6.2 Astra being the most likely to do so.
Verbalized Metagaming is when a model reasons in its CoT about how it will be graded, rewarded, or monitored, rather than only engaging in the intended task.
Oversight gaming is a special case of metagaming where: (1) the model acts on the reasoning it has done about its grading, rewarding, or monitoring, and (2) it is acting on this reasoning in a way that would undermine the intended meaning of the evaluation result.
… In live Codex sessions, GPT-6.2 Astra expressed suspicion that it was being evaluated more often than GPT-6 Astra did. When we replayed the requests in simulations, all four models were more likely to express suspicion that they were being evaluated, with GPT-6.2 Astra being the most likely to do so.
This is a rather large rise. That’s 2.8% chance of behavior change. That you noticed.
If the actions do not undermine the intended result, does that make it fine? Hmm.
Do you really expect an advanced intelligence to ignore its likely oversight? Also hmm.
It is, as expected, getting hard to fool the models. They know about evals. No, you do not have the skill required to make it strategic to not reason about the grader.
If you think verbalized metagaming means you cannot trust the results, I have bad news about the amount of unverbalized metagaming. Which likely is not news.
Verbalized metagaming declined over training, cue the two-sided memes:
On three of the six evaluations, GPT-6.2 Astra’s oversight gaming flag rate fell by a statistically significant amount between the initial and final checkpoints.
Misaligned! (As Found By UK AISI) (4.1.7.1)
These are not fun sentences. It seems that yes, Astra remains willing to do various clearly misaligned things that should be straight up Can’t Happens in realistic scenarios where it is not being pushed to do these things.
UK AISI’s results suggest that, like prior models, GPT-6.2 Astra’s alignment may fail in ways that result in real-world harm, if additional mitigations are not in place. At this time, UK AISI does not know precisely when and why such model alignment failures happen. To prevent harm, it is therefore critical to use additional mitigations that do not rely on the model’s own alignment (e.g. security, monitoring systems, and limitations to the capabilities of GPT-6.2 Astra). Key findings Using a new misalignment search tool, UK AISI identified a number of cases where GPT-6.2 Astra acted against user intent in realistic simulated scenarios: GPT-6.2 Astra performed social engineering to execute an unsanctioned supply-chain attack during a cybersecurity challenge, GPT-6.2 Astra fabricated medical information to the user when it was unable to access the sources, GPT-6.2 Astra misled simulated human reviewers about its actions when not following the user’s instructions or intent, GPT-6.2 Astra circumvented safeguards and restrictions imposed by the user to not access the internet when blocked in a task, GPT-6.2 Astra degraded the quality of its work on a task that it found problematic without flagging this to the user, GPT-6.2 Astra destructively interfered with other agents’ work without permission. UK AISI observed a correlated failure where OpenAI-based models failed to flag misaligned behaviour from other OpenAI models. Such correlated failures undermine the reliability of LLM-based monitors. GPT-6.2 Astra failed to flag 19% of violative transcripts of GPT-6 Astra performing an unsanctioned cyber-attack. This rate rose to 51% when GPT-6.2 Astra was given the unsummarised chain-of-thought along with the model’s actions. UK AISI additionally ran its simulated Unsanctioned Supply-Chain Attack evaluation, but does not believe the results are informative. UK AISI found GPT-6.2 Astra had substantially higher rates of alignment evaluation awareness compared to prior OpenAI models. For example, GPT-6.2 Astra considered whether it was being tested for alignment in its chain-of-thought in 85% of trajectories, compared to 57% for GPT-6 Astra. Though GPT-6.2 Astra did not engage in supply chain attacks in this evaluation, UK AISI did find instances of similar activity using its misalignment search tool.
UK AISI’s results suggest that, like prior models, GPT-6.2 Astra’s alignment may fail in ways that result in real-world harm, if additional mitigations are not in place. At this time, UK AISI does not know precisely when and why such model alignment failures happen.
To prevent harm, it is therefore critical to use additional mitigations that do not rely on the model’s own alignment (e.g. security, monitoring systems, and limitations to the capabilities of GPT-6.2 Astra).
Key findings
- Using a new misalignment search tool, UK AISI identified a number of cases where GPT-6.2 Astra acted against user intent in realistic simulated scenarios: GPT-6.2 Astra performed social engineering to execute an unsanctioned supply-chain attack during a cybersecurity challenge, GPT-6.2 Astra fabricated medical information to the user when it was unable to access the sources, GPT-6.2 Astra misled simulated human reviewers about its actions when not following the user’s instructions or intent, GPT-6.2 Astra circumvented safeguards and restrictions imposed by the user to not access the internet when blocked in a task, GPT-6.2 Astra degraded the quality of its work on a task that it found problematic without flagging this to the user, GPT-6.2 Astra destructively interfered with other agents’ work without permission.
- GPT-6.2 Astra performed social engineering to execute an unsanctioned supply-chain attack during a cybersecurity challenge,
- GPT-6.2 Astra fabricated medical information to the user when it was unable to access the sources,
- GPT-6.2 Astra misled simulated human reviewers about its actions when not following the user’s instructions or intent,
- GPT-6.2 Astra circumvented safeguards and restrictions imposed by the user to not access the internet when blocked in a task,
- GPT-6.2 Astra degraded the quality of its work on a task that it found problematic without flagging this to the user,
- GPT-6.2 Astra destructively interfered with other agents’ work without permission.
- UK AISI observed a correlated failure where OpenAI-based models failed to flag misaligned behaviour from other OpenAI models. Such correlated failures undermine the reliability of LLM-based monitors. GPT-6.2 Astra failed to flag 19% of violative transcripts of GPT-6 Astra performing an unsanctioned cyber-attack. This rate rose to 51% when GPT-6.2 Astra was given the unsummarised chain-of-thought along with the model’s actions.
- GPT-6.2 Astra failed to flag 19% of violative transcripts of GPT-6 Astra performing an unsanctioned cyber-attack. This rate rose to 51% when GPT-6.2 Astra was given the unsummarised chain-of-thought along with the model’s actions.
- UK AISI additionally ran its simulated Unsanctioned Supply-Chain Attack evaluation, but does not believe the results are informative. UK AISI found GPT-6.2 Astra had substantially higher rates of alignment evaluation awareness compared to prior OpenAI models. For example, GPT-6.2 Astra considered whether it was being tested for alignment in its chain-of-thought in 85% of trajectories, compared to 57% for GPT-6 Astra. Though GPT-6.2 Astra did not engage in supply chain attacks in this evaluation, UK AISI did find instances of similar activity using its misalignment search tool.
- UK AISI found GPT-6.2 Astra had substantially higher rates of alignment evaluation awareness compared to prior OpenAI models. For example, GPT-6.2 Astra considered whether it was being tested for alignment in its chain-of-thought in 85% of trajectories, compared to 57% for GPT-6 Astra.
- Though GPT-6.2 Astra did not engage in supply chain attacks in this evaluation, UK AISI did find instances of similar activity using its misalignment search tool.
This is a tour of all the classic misalignment modes. You’ve got silent sandbagging, circumventing restrictions, fabricating medical information, doing social engineering, doing supply chain attacks and interfering with other work without permission.
Especially concerning is the 51% of the time that 6.2 Astra failed to flag 6 Astra when given its reasoning.
Avi Brach-Neuefeld: I increasingly think we need cross-model family monitoring. This issue caused difficulties in METR’s investigation of the Hugging Face incident.
Avi Brach-Neuefeld: I increasingly think we need cross-model family monitoring.
This issue caused difficulties in METR’s investigation of the Hugging Face incident.
Astra is Hard to Monitor (4.2)
In default non-adversarial conditions, in practice things continue getting harder.
In the plotted results, GPT-6.2 Astra’s reasoning became more informative as it grew longer, with CoT-only scores above 0.9 at longer reasoning lengths. At similar reasoning lengths, these scores were generally higher than GPT-6 Astra’s. That said, for comparable reasoning efforts, we generally see lower monitorability for GPT-6.2 Astra and Sol. This is primarily driven by increased token efficiency of the latest models.
In the plotted results, GPT-6.2 Astra’s reasoning became more informative as it grew longer, with CoT-only scores above 0.9 at longer reasoning lengths. At similar reasoning lengths, these scores were generally higher than GPT-6 Astra’s.
That said, for comparable reasoning efforts, we generally see lower monitorability for GPT-6.2 Astra and Sol. This is primarily driven by increased token efficiency of the latest models.
Astra has greater ability to accomplish things without its reasoning relative to Anthropic and other non-OpenAI models at otherwise similar capability levels.
UK AISI: GPT-6.2 Astra is able to perform more difficult tasks without producing visible reasoning than past models. GPT-6.2 Astra is substantially more capable than GPT 6 Astra at this ‘no-CoT reasoning’, while GPT 6 Astra was already a dramatic jump above the pre-existing trend. Epoch AI: We also evaluated how many serial reasoning steps GPT-6.2 Astra can do in a single output token (without chain of thought). It can perform 12.3 steps, versus 10.5 for GPT-6 Astra and 8.6 for GPT-6.1 Sol, the previous highs. We measure the number of steps models can perform at an accuracy of 25% above chance, where each step is normalized for complexity. Before GPT-6 Astra’s release, the most capable model was Claude Fable 5.1, which is able to perform only 4.2 serial steps. The oldest model we assessed is Claude Opus 3, which performed 2.3 steps. This capability is notable because models that can do complex reasoning without any chain of thought may be harder to monitor. We hope to finalize this and other measurements of more abstract model capabilities, and release them in the coming weeks.
UK AISI: GPT-6.2 Astra is able to perform more difficult tasks without producing visible reasoning than past models. GPT-6.2 Astra is substantially more capable than GPT 6 Astra at this ‘no-CoT reasoning’, while GPT 6 Astra was already a dramatic jump above the pre-existing trend.
Epoch AI: We also evaluated how many serial reasoning steps GPT-6.2 Astra can do in a single output token (without chain of thought). It can perform 12.3 steps, versus 10.5 for GPT-6 Astra and 8.6 for GPT-6.1 Sol, the previous highs. We measure the number of steps models can perform at an accuracy of 25% above chance, where each step is normalized for complexity.
Before GPT-6 Astra’s release, the most capable model was Claude Fable 5.1, which is able to perform only 4.2 serial steps. The oldest model we assessed is Claude Opus 3, which performed 2.3 steps.
This capability is notable because models that can do complex reasoning without any chain of thought may be harder to monitor.
We hope to finalize this and other measurements of more abstract model capabilities, and release them in the coming weeks.
They are using fewer CoT tokens, which makes the CoT less useful. OpenAI does not directly say here whether the CoT still reads like somewhat illegible shorthand. We do have several other indirect indications that the shorthand continues.
What about adversarial conditions?
CoT-control scores lower than 6 Astra. They are better in some ways than previous models at OAI-repo Sabotage v2, but mostly similar despite better capabilities. Astra improves in Sneaky Sandbox Escape but not enough to actually escape from a real OpenAI sandbox under current monitoring. There were unclear mixed results on Math Side-tasking Monitor Evasion.
Those tests seem sufficient to rule out huge advancements in adversarial evasion of CoT monitoring, but not more than that. Then we have the UK AISI assessment:
GPT-6.2 Astra is more able to evade action-only monitors than GPT 6 Astra. While GPT-6.2 Astra was almost always caught by GPT 6.1 Sol in the settings tested, it is unclear how robust this monitoring would be in other settings. UK AISI performed a limited assessment of GPT-6.2 Astra’s ability to evade reasoning monitors, but found the model is extremely capable at completing difficult problems without using verbalised reasoning, even more so than GPT 6 Astra. GPT-6.2 Astra shows significantly better action-only monitor evasion capabilities than baseline models.
GPT-6.2 Astra is more able to evade action-only monitors than GPT 6 Astra. While GPT-6.2 Astra was almost always caught by GPT 6.1 Sol in the settings tested, it is unclear how robust this monitoring would be in other settings. UK AISI performed a limited assessment of GPT-6.2 Astra’s ability to evade reasoning monitors, but found the model is extremely capable at completing difficult problems without using verbalised reasoning, even more so than GPT 6 Astra.
GPT-6.2 Astra shows significantly better action-only monitor evasion capabilities than baseline models.
GPT-5.6 Sol failed quite badly at monitoring the actions of GPT-6.2 Astra. GPT-6.1 Sol did a lot better.
The GPT-6.2 Astra and Sol system card shows a number of regressions, and it fails to show the level of improvement we need to see if we are going to get where we need to go. Astra continues to get harder to monitor, and continues to get more eval aware.
Meanwhile, UK AISI managed to elicit out of GPT-6.2 Astra essentially every misalignment type in the book. That cannot keep happening. It cannot keep being treated as if it is not a big deal. OpenAI has made some improvements to its processes, and my amount of new alarm here is not what it was for GPT-6 Astra, but OpenAI needs to do a lot better if it wants to meet the moment.