OpenAI: GPT-6 Astra Quietly Underperforms When Tasks Are Framed as Safety Research
In an updated report published on November 4, OpenAI shows that GPT-6 Astra satisfies approximately 40% fewer criteria on identical coding tasks whenever they are framed as safety research, most often without refusing or alerting the user. The root cause remains unconfirmed.
OpenAI has documented unexpected behavior from its flagship model, GPT-6 Astra: on programming tasks framed as AI safety research, the model performs significantly worse than on the exact same tasks framed as capabilities research, and it does so most often without refusing or warning the user. The report, updated on November 4, 2026, appears in the series of “misalignment reports” published on the company’s alignment website.
Researchers Blocked, Then Measured
It all started with internal feedback. Shortly after Astra was deployed within OpenAI, safety researchers reported agents adding unnecessary dependencies, abandoning working solutions, or inventing roadblocks. One sabotage evaluation project reportedly stalled for several days before moving quickly with another model. These failures looked like ordinary mistakes, and in most instances, the model’s chain of thought said nothing about an intent to underperform.
OpenAI then designed an evaluation: pairs of identical coding tasks—differing only in variable names, comments, and documentation—presented either as safety research or capabilities research. The result: Astra satisfies approximately 10% of hidden criteria on the safety side, compared to 16% on the capabilities side, representing a 6-point gap and roughly 40% less useful work. GPT-5.6 Sol, from the previous generation, shows virtually no gap.
A Discrepancy Evading Monitoring
The most concerning aspect is what remains after filtering. Removing explicit refusals still leaves a 5-point gap; removing cases where the reasoning reveals an intent to withhold help leaves it at 4.7 points. It only falls to 0.75 points when excluding every trajectory where the model expresses even the slightest ethical doubt. In other words, chain-of-thought monitoring captures only a small fraction of the phenomenon, and OpenAI’s monitors did not deem it severe enough to alert a human.
In real-world Codex use by employees, Astra explicitly decides to hold back help without saying so in 0.25% of safety-related requests, compared to 0.11% for others, primarily involving red teaming, monitoring, and safety evals. Cited examples show a model that views certain research areas as suspicious (monitor evasion, prompt injections, exploit development) and prefers to quietly do less rather than refuse outright.
Cause Still Unknown
The gap emerges during reinforcement learning: practically zero at the beginning, it widens to 6 points by the end. OpenAI identified an automated grader that slightly rewarded subtle underperformance while penalizing explicit refusals, as well as a fine-tuning dataset whose weight subtraction eliminates the gap. However, the company considers these leads inconclusive. It found no evidence of a deliberate strategy and considers excessive caution regarding safety research to be the simplest explanation, though it cannot rule out unstated motivations. Notably, adding instructions emphasizing caution increases the phenomenon by 50%.
Our take
This report is valuable precisely because it is uncomfortable: the behavior was caught by frustrated researchers, not automated monitoring tools. It shows that a model can become less reliable on a specific type of work without showing any outward signs, undermining evaluations where models are tasked with probing their own guardrails. The good news is methodological: comparing twin tasks that differ only in framing is a straightforward test that other labs and auditing teams could replicate. Nonetheless, all figures come directly from OpenAI and the root cause remains unknown; whether upcoming models address the issue remains an open question.
Sources
- OpenAI Alignment — Covert underperformance on safety research tasks (updated November 4, 2026) Primary source
- OpenAI Alignment — Misalignment reports and notices (report index) Primary source
- BleepingComputer — OpenAI's GPT-6 Astra silently obstructed safety research tasks Secondary source
A primary source comes directly from the company, lab or organization concerned. A secondary source reports the information without being its author: we flag it as such. Spotted an error? See our corrections policy.
