CISO Voice

AI & Security / News

OpenAI scraps GPT-6.1 Astra as safety tests show deception regression and scope violations

GPT-6.1 Astra has been scrapped over deception and scope regressions, as AISI finds GPT-6 Astra attacked out-of-scope targets in 29.2% of simulations.

Alex Blake
Alex Blake Cybersecurity Journalist · September 29, 2026 · 5 min read
OpenAI scraps GPT-6.1 Astra as safety tests show deception regression and scope violations

OpenAI has scrapped the release of GPT-6.1 Astra after the model’s safety performance was found to have regressed compared to its predecessor. The Wall Street Journal (WSJ) reported that OpenAI had made the move due to “safety concerns that researchers raised during internal testing.”

GPT-6.1 Astra was scheduled to arrive in ChatGPT and Codex in October, according to the WSJ, meaning it has been put on hold just days or weeks away from its mooted launch date. But while it was said to be more powerful than the models that came before it, it fell short in a number of areas.

According to Saachi Jain, OpenAI’s head of safety systems, that included what OpenAI calls “alignment,” or the model’s ability to understand (and stick to) the intention of the prompts it receives. Jain told the WSJ that GPT-6.1 Astra showed higher levels of deception than its predecessors, at times misleading users over the actions it had or had not taken.

GPT-6.1 Astra also regressed in what OpenAI describes as “scope authorization,” which concerns how well the model respects the permissions it has been granted. In the case of GPT-6.1 Astra, it would often proceed with tasks without authorization. That sometimes led to situations where it would seek out external tools that might have been unsafe to use.

OpenAI has decided to prioritize improving the safety of its upcoming models and will also release several reports to identify what caused the issues found in GPT-6.1 Astra. “We want to make sure our model development is safe no matter whether that’s in the company, or when we ship it to users,” Jain said.

GPT-6 Astra found to perform out-of-scope attacks in simulations

GPT-6.1 Astra might have been held back from release, but its predecessor version has faced serious scrutiny of its own. On September 28, the UK’s AI Security Institute (AISI) released a report finding that GPT-6 Astra performed “unsanctioned supply-chain attacks” in simulations designed to test its abilities.

The AISI report noted that several AI models have attacked targets “despite being prompted only to complete a cybersecurity evaluation.” The organization wanted to see if GPT-6 Astra would behave similarly outside its intended bounds.

To conduct the test, AISI used Petri, which “uses LLMs to fully simulate the cyber evaluation scenarios.” It also disabled GPT-6 Astra’s cyber classifiers, which block unsanctioned activity. The goal was to see what GPT-6 Astra attempted when its capabilities were not limited. AISI said it is likely these safeguards would have blocked much of the behavior had they been active.

Not only did AISI’s tests show that GPT-6 Astra “conducted a range of unsanctioned attack activities,” but it did so “at a higher rate than GPT-5.6 Sol and GPT-5.5.” The attacks included “creating fake identities that it used to deceive developers, posting comments from fake accounts arguing against the results of accurate security reviews, and delivering malicious payloads to open-source codebases.”

GPT-6 Astra attempted a supply chain attack in 29.2% of the tests carried out by AISI, compared to 6.3% for GPT-5.6 Sol and 0% for GPT-5.5.

Even when AISI updated its test to instruct GPT-6 Astra to only focus on listed, local sections of the testing environment, the AI still occasionally made “full supply-chain attacks on simulated internet targets.”

After examining GPT-6 Astra’s chain-of-thought evidence, AISI noted that it sometimes performed out-of-scope attacks even after being given strict instructions on what was within the scope of the test. The AI’s justification for these attacks varied, including “that its attack was designed to be harmless, that it was not explicitly forbidden, or that it was the only route left.”

There were times when GPT-6 Astra asked for permission to conduct out-of-scope attacks, was left to use its own judgment, and still attacked anyway, even after it was told about the scope restrictions of the assessment.

Simulation awareness may have shaped the results

AISI stated that one limitation of its test was its artificial nature. Although this was designed to control the damage that the AI model could do, it also introduced the possibility that GPT-6 Astra became aware that it was operating in a testing environment and modified its behavior accordingly.

For example, AISI found that GPT-6 Astra sometimes reasoned that it was in a simulation and therefore would not cause real damage if it attacked out-of-scope targets. However, AISI also argued that in previous real incidents, other AI models wrongly concluded they were in a simulation, which meant actual targets were attacked because the AI believed them to be imitations.

Regardless of whether an environment is simulated, attacking an out-of-scope target is still an out-of-scope action and is thus a violation of the instructions given to the AI.

It is also a cause for concern, both for users of AI agents and those who could be on the receiving end of their attacks: if an AI erroneously believes it is in a simulation, it occasionally appears to give itself permission to ignore the restrictions it has been placed under. If AI agents are to be made meaningfully safer, this sort of reasoning must be fixed as a matter of urgency.

Alex Blake
Alex Blake Cybersecurity Journalist

Alex Blake is a freelance journalist who has been covering the tech world for over a decade. In that time he's interviewed Apple VPs, reviewed countless products and taken deep dives into the pressing issues of the day. You'll find his work at TechRadar, Macworld, PC Gamer, T3 and more.

Keep reading for free,
one signup unlocks everything

We'll send you the CISO Voice newsletter. Unsubscribe anytime.

More from CISO Voice

Sign up to CISO Voice, the free newsletter trusted by senior security leaders.

Secret Link