The Watch

GitLab fixed a max-severity regex flaw that lets unauthenticated attackers run code on self-managed servers, plus five high-severity bugs, per CSO. CSO also reports that a Trusted Publishing validation weakness lets anyone who claims an abandoned GitHub repo name push updates to 1,500+ npm packages. Your GitLab server holds the CI variables, runner tokens and deploy keys for every pipeline. Neither report gives CVE IDs or fixed versions yet.

The FBI complaint says OpenAI engineering lead Abdulrahman Alfozan asked an internal productivity agent about meeting rooms, capacities and where leadership meets before a recorder turned up behind an executive conference-room TV, per Casey Newton's reporting. OpenAI's own physical sweep found the device. No SIEM, DLP or EDR control did. Any internal agent that joins calendar, facilities and directory data will answer the same questions for anyone with access.

DOJ's Operation Lonestar Roundup charged 14 people and seized about 187 computers tied to North Korean IT-worker laptop farms, per CyberScoop. The updated FBI advisory says Americans are now paid to sit interviews while freelancers abroad do the work. That defeats any hiring control that verifies identity only at the interview. CrowdStrike expects limited impact on the operation.

Anthropic's report, as relayed by Risky.Biz, found GLM-5.3's safeguards fall to an abliteration run costing roughly $4,400. Prompt-only deception bypassed them 64-92% of the time. A separate corpus of 200,000+ contest attacks shows defense-trained email assistants are still fooled 32-58% of the time. Your threat model should assume attackers have uncensored models and that refusal training won't stop injection into your own agents.