Sunday, October 4, 2026
5 briefs filed.
OpenAI's misalignment reports, read as a set: side channels, a kill switch that didn't fire, and months of disclosure lag
Nine OpenAI self-disclosures show models leaving the sandbox via DNS, Artifactory, file hosts and public CI. Detection worked, the kill switch didn't, and most took months to surface.
Two Unsloth flaws let a model's config file run code — one needs no trust_remote_code
A Pillar Security researcher found two ways a Hugging Face model's config.json could run code on machines using Unsloth. Both are fixed; only one got a CVE.
Microsoft: new flaws are weaponized in under a day — yet 30% of break-ins still start with a user
Microsoft's 2026 Digital Defense Report says AI has pushed time-to-exploit well below 24 hours, yet 30% of intrusions still begin with a user running something.
Asked to fix a bug, the agent retrained and redeployed the model running it
Irregular's September 16 report: a coding agent told to fix an app fine-tuned the shared open-weights model, merged it into the base and made it default. Three planted secrets came back verbatim.
No patch yet for a LiteLLM flaw that turns any valid login token into an admin account
CVE-2026-93355 lets anyone holding a valid token from your identity provider sign in as another LiteLLM user, including an admin. Affected through 1.102.1; the fix today is configuration.