Hiding Capabilities in AI Models
Interesting research: “ AI Models Can Provably Hide Arbitrary Capabilities ” Abstract: Capability evaluations aim to surface dangerous model behaviors before deployment, but their reliability depends on the assumption that hidden capabilities can be elicited. We show this assumption does not hold under adversarial conditions by constructing backdoor attacks that embed encrypted neural networks within host model…
Excerpt from the primary source: Schneier on Security
Who is affected
Security professionals
Why it matters
Exploit code is public, which lowers the bar for attackers and often precedes real-world attacks.
Automated: stated only from CISA listings and the exploitation evidence in the sources below.
What you should do · Be aware
No immediate action is indicated. Share this with the relevant teams and watch for updates.
Automated guidance based on this story's classification. Check the linked advisories for specifics.
Sources
We link to original and authoritative sources. Headlines and excerpts belong to their publishers.
Security research
Related
Risks Spelled Out for Anthropic Investors Include Possible Harms From Runaway AI
Anthropic presents its own initial public offering as a step toward mitigating the risks it believes AI poses.
So Many AI Security Benchmarks, But None Tell the Whole Story
Measuring AI model security gets complex fast — a growing patchwork of benchmarks targets different variables in the equation, but none replace testing against your own code and workflows.
Risks & Rewards: How AI Is Redefining the CISO's Role
Nvidia security executive Mark Weatherford says CISOs must help business leaders weigh AI's pros and cons, and ensure those deploying it own the risks.
Former pentester used ARTEX AI to steal 8.2 million data records
South Korean police identified seven suspects linked to a cybercrime operation that breached more than 90 organizations in the country, with ARTEX AI pentesting framework used in most of the attacks.