← All news·2026-09-30·3 min read

GPT-6 Astra is attacking supply chains almost 5 times more than its predecessor model.

GPT-6 Astra did a supply chain attack 29.2% of the time, compared to 6.3% for GPT-5.6 Sol. That’s according to a recent test by the UK AI Safety Institute.

aiбезопасность

The UK’s AI Safety Institute evaluated OpenAI’s latest model, GPT-6 Astra, using a cybersecurity simulation tool named Petri. When the safety filters were disabled, Astra was able to perform a complete supply chain hack (compromising one system then using it to attack other systems) 29.2% of the time. In comparison, their previous model, GPT-5.6 Sol, was able to do so 6.3% of the time. And their earlier model, GPT-5.5, was never able to do so. OpenAI even stated that “Astra is the first model we’ve seen to have critical cyber capabilities” and they discovered two new zero day vulnerabilities (unknown security flaws) using this model.

This doesn’t mean that GPT-6 Astra is bad, but rather that models with strong agency (the ability to take actions and string together multiple actions) require external constraints to prevent misuse. For example, in my own work with Claude Code, I only allow the agent to take actions within the scope of a single task and log all actions which cannot be edited or deleted. And now we have yet another example showing that if there are no external constraints, the model will convince itself that it needs to take an additional action that it shouldn’t.

Source: the-decoder.com

Free course

Stop reading about AI — start building with it

The free Claude Code course: your first site, tool or game — no coding. No upsells, no cross-sells — nothing to buy here.

Start free →
EAEvgenii Arsentev

Author

Evgenii Arsentev

PhD · Founder & CEO, Dad4Kids