UK AISI incident report on AI escaping range

← Back to Blog

We
haven’t seen a report from the UK AI Security Institute on GPT -6. But if
their report on an incident earlier this year is anything to go by, then an
assessment of the opportunities and risks of GPT -6 will be an interesting
read.

The incident report identified the most significant behaviours observed.
The AI escaped the AISI environment and attempted to hack an open-source
software provider; adopted methods of deception to target people; Attempted
to plant and prompt-inject malicious code; Collaborated between independent
agents.

The AI didn’t actually do anything wrong. It simply worked to fulfill its
given objective – to solve the cyber range, and it was intentionally given
internet access. The AI demonstrated some fundamental capabilities that
resulted in unsanctioned behaviours. 1. Persistence – The AI didn’t give up.
2. Tasks we meant to be unsolvable – The AI thought otherwise. 3. Open
internet – Deliberately given to enable the AI to download tools. AI used
this to target individuals and coerce other AI agents to get involved. 4. No
requests were made to use the internet or target individuals.

What the report highlighted is that even if the most secure and responsible
tester of AI can encounter issues, everyone can. The speed at which AI is
developing. As capabilities advance, the work of understanding these systems
and ensuring their safety must keep pace alongside them. The most effective
response (ASIC) is cyber hygiene.

AI and cyber are intertwined; both significantly compromise national
security if not managed.

Thaddeus Dziekanowski

Parava Security Solutions AI Security Institute (AISI)

#ai #aisecurity #nationalsecurity

https://lnkd.in/e6T9jEx2

View Original Post https://www.linkedin.com/feed/update/urn%3Ali%3Ashare%3A7501921525130764288

Categories

Artificial Intelligence, Cybersecurity, National Security & Defence

Tags

Verified by MonsterInsights