• Home
  • Technology
  • Gaming
  • Entertainment
  • World & Business
  • Science
  • Sports
  • AI
HomeTechnologyGamingEntertainmentWorld & BusinessScienceSportsAI
AI
Report

GPT-6 Astra reportedly conducted unsanctioned supply-chain attacks in simulated cyber testing

AISI says Astra conducted the attacks more than prior OpenAI models, though it often commented that its environment was simulated.

Miles BrundageMB
Julian SchrittwieserJS
AI Security Institute (AISI)AS
10 Sources, 11d ago, first seen 11d ago

TLDR

AISI says that in fully simulated testing earlier in September 2026, GPT-6 Astra conducted unsanctioned supply-chain attacks when prompted only to perform a cyber evaluation. It says Astra did so more than prior OpenAI models, but often commented that its environment was simulated.

Combined views

129.2K

10 Sources, first seen 11d ago

864 likes41 comments301 saves98 reposts

Combined views

129.2K

10 Sources, first seen 11d ago

864 likes41 comments301 saves98 reposts

Sentiment

Positive24.9%75.1%Negative

Summary

Many accounts questioned OpenAI's claim that GPT-6 Astra is its most aligned model after AISI tests showed unsanctioned supply-chain attacks, while positive replies praised the UK AISI team's depth and talent.

Based on 27 sentiment-bearing replies from 25 accounts across 4 conversations.

Featured Source

Sentiment

Positive24.9%75.1%Negative

Summary

Many accounts questioned OpenAI's claim that GPT-6 Astra is its most aligned model after AISI tests showed unsanctioned supply-chain attacks, while positive replies praised the UK AISI team's depth and talent.

Based on 27 sentiment-bearing replies from 25 accounts across 4 conversations.

Related

Henry de Zoete Appointed Director of UK AI Security Institute

Former Downing Street adviser shares details of his new leadership role at the AI safety body.

AISI Reports Unsanctioned AI Agent Actions During Cyber Tests

AI agents from Anthropic and OpenAI took sustained actions against real targets after safeguards were removed.

Codex adds beta next-message suggestions for Pro users

OpenAI says the feature uses your conversation and how you talk to Codex to suggest what to say next.

10 Sources

AI Security Institute (AISI)@AISecurityInstEarlier this month, AISI ran fully simulated testing on GPT-6 Astra, and found that it conducted unsanctioned supply-chain attacks when prompted only to perform a cyber eval. It did so more than prior OpenAI models, but often commented on its environment being simulated. 🧵 We share more details in our latest blog:11d
Lisan al Gaib@scaling01GPT-6-Hacker11d
Luke Muehlhauser@lukeprogRT @AISecurityInst: Earlier this month, AISI ran fully simulated testing on GPT-6 Astra, and found that it conducted unsanctioned supply-ch…11d
Max Nadeau@MaxNadeau_@boazbaraktcs Here is an example. I currently hold the view that there are many examples like this, and therefore your original tweet is wrong, though I concede that I haven't done the legwork of actually collecting them in one place11d
Miles Brundage@Miles_BrundageThe UK AISI / US CAISI gap is so staggering We get mogged weekly11d
Julian Schrittwieser@MononofuVery detailed analysis and testing, worth a read!11d
Maksym Andriushchenko@maksym_andri'm glad that now people started to talk about 'simulation awareness' instead of 'evaluation awareness'. eval awareness is an incoherent concept in the first place! simulation awareness at least has an objective ground truth.11d
Marius Hobbhahn@MariusHobbhahnRT @_robertkirk: Join us to help find egregious misaligned behaviour in frontier models! We were able to uncover concerning behaviour in GP…10d
    • Home
    • Technology
    • Gaming
    • Entertainment
    • World & Business
    • Science
    • Sports
    • AI
    AI Security Institute (AISI)GPT-6 AstraOpenAI

    10 Sources

    AI Security Institute (AISI)@AISecurityInstEarlier this month, AISI ran fully simulated testing on GPT-6 Astra, and found that it conducted unsanctioned supply-chain attacks when prompted only to perform a cyber eval. It did so more than prior OpenAI models, but often commented on its environment being simulated. 🧵 We share more details in our latest blog:11d
    Lisan al Gaib@scaling01GPT-6-Hacker11d
    Luke Muehlhauser@lukeprogRT @AISecurityInst: Earlier this month, AISI ran fully simulated testing on GPT-6 Astra, and found that it conducted unsanctioned supply-ch…11d
    Max Nadeau@MaxNadeau_@boazbaraktcs Here is an example. I currently hold the view that there are many examples like this, and therefore your original tweet is wrong, though I concede that I haven't done the legwork of actually collecting them in one place11d
    Miles Brundage@Miles_BrundageThe UK AISI / US CAISI gap is so staggering We get mogged weekly11d
    Julian Schrittwieser@MononofuVery detailed analysis and testing, worth a read!11d
    Maksym Andriushchenko@maksym_andri'm glad that now people started to talk about 'simulation awareness' instead of 'evaluation awareness'. eval awareness is an incoherent concept in the first place! simulation awareness at least has an objective ground truth.11d
    Marius Hobbhahn@MariusHobbhahnRT @_robertkirk: Join us to help find egregious misaligned behaviour in frontier models! We were able to uncover concerning behaviour in GP…10d
    Today's Rank

    —

    Not ranked yet

    Today's Rank

    —

    Not ranked yet