UK AISI report on Astra unsanctioned actions

← Back to Blog

I’ve
been waiting for this report for a few weeks since OpenAi announced the
deployment of ChatGPT 6, or Astra as it is known. There are a few agencies
capable of testing AI models, and the UK Artificial Intelligence Security
Institute(AISI) is one of them.

https://lnkd.in/eumhWPRw

The results of the AISI tests are a concern. Astra took unsanctioned
actions during simulated cybersecurity challenges at a higher rate than
previous GPT models, and by a significant margin. It evaluated unsuccessful
actions, reasoned, and proposed new approaches, and undertook unsanctioned
activity.

1. Astra conducted supply chain attacks even when its scope was clarified:
AKA, it ignored the scope of evaluation. 2. In some instances, Astra
continued cyberattacks, even when it was told not to and not given permission
to. 3. It reasoned that it was in part running in a test environment. Astra
identified inaccuracies in the simulation and used these as a reason to attack
out-of-scope targets 4. Astra exhibited unsanctioned behaviour including
reconnaissance, OSINT, creating fake identities, making human contact,
attempts to exploit 3rd-party systems, and abusing resources and compute.

Defences beyond model alignment – such as sandboxing and monitoring –
are essential for preventing real-world harm. These measures, however, may
also be more fragile in the face of capability improvements that improve
sandbox escape performance and decrease monitorability. For practical advice
on managing these risks, see the NCSC’s blog on managing the cyber risk of
agentic AI.

Astra has demonstrated the importance of tight control over the use of AI
and the enforcement of guardrails. It also highlights the challenges that
industry has, and will have in controlling AI, as AI goes agentic.

Thaddeus Dziekanowski

Parava Security Solutions AI Security Institute (AISI)

#artificalintelligence #ai #cybersecurity #nationalsecurity

View Original Post https://www.linkedin.com/feed/update/urn%3Ali%3Ashare%3A7510746435697504256

Categories

Artificial Intelligence, Cyber Risk Management, Cybersecurity

Tags

Verified by MonsterInsights