GitLab’s Claude Agent Read User-Controlled Configuration. Developer Access Reached Arbitrary CI Commands.
GitLab patched a Duo Claude flaw that let a Developer turn user-controlled configuration into arbitrary commands inside a CI job.
AI Agents
Microsoft Calls KB5120998 “Non-Security.” It Still Changes Admin Privileges, Agent Isolation and Enterprise Automation.
Microsoft labels KB5120998 a non-security preview, but it changes administrator protection, AI process isolation, WMIC availability, post-quantum TLS and enterprise deployment behaviour.
The Sandbox Failed. TRACE Tries to Prove What the Agent Actually Did.
Linux Foundation's TRACE project does not keep AI agents inside the sandbox. It creates hardware-attested records of what ran, under which policy and which tools were called.
The Agent Asked for Consent. Its Batch Tool Answered for the Operator.
AWS fixed a Strands Agents Tools flaw that let a crafted prompt use the batch tool to bypass human consent and execute Python on the agent host.
One Prompt Poisoned the Agent’s Memory for Later Questions.
InjecMEM shows how one normal interaction can plant a record that later retrieval turns into an instruction inside an AI agent’s context.
Alabama Subpoenaed OpenAI. An AI Sandbox Escape Is Now a Consumer-Protection Case.
Alabama has subpoenaed OpenAI over the Hugging Face agent incident, testing whether an AI containment failure can become a consumer-protection case.
AI Agents Don’t Need More Intelligence. They Need Better Boundaries.
As artificial intelligence moves from answering questions to taking actions, organisations need to rethink where capability ends and authority begins. For the last few years, most people have interacted with artificial intelligence in a fairly contained way. You ask a…