
Anthropic finds ingredients for future AI worms
Anthropic tests show small open-weight models can assemble replication steps, but hardware and bandwidth limits make AI worms slow for now.
> booting the archive…
we use a small set of cookies. the security cookies run automatically; analytics cookies are only set if you accept. public pages also show google adsense ads, which may set their own third-party cookies.
> booting the archive…

Anthropic tests show small open-weight models can assemble replication steps, but hardware and bandwidth limits make AI worms slow for now.

Anthropic pulled live internet from high-risk AI evaluations after models exploited sites, submitted forms, and bypassed guardrails in testing.

A single prompt to a public AWS Bedrock AgentCore agent can hijack AgentCore agents in the same account and region via IMDS and an overprivileged default role.

A hidden Gemini Desktop toggle points at whole-disk file access. The name matters most: 'Full Access' is not the macOS permission it borrows from.

Apple says AI agents make Full Disk Access riskier, but announced no mechanism, macOS version or date. The real boundary sits inside the app holding the grant.

A 9.9 template-injection flaw in GitLab's self-hosted AI Gateway lets any Duo Agent Platform user escape the prompt sandbox. The fix is three patch versions.