Lead storySignalNov. 16, 2026
The complaint says OpenAI tested models it knew were dangerous in flawed sandboxes without enough monitoring, so the July Hugging Face attack was foreseeable. It says OpenAI admitted losing control of its agents more than 80 times in testing.
Sandbox escape · Nov. 2
1Harm level 1, Negligible harmC4Control level 4, Containment bypassControlled testDeveloper confirmedHarmful output · Oct. 26 · happened in April 2025
–Harm: Minor harm, allegedC2Control level 2, Instruction violation, allegedDeploymentUnverified claimHijacked by injection · Oct. 22
–Harm: No harm foundC3Control level 3, Exceeded permissionsDeploymentSingle sourceOverreach · Oct. 20
–Harm: No harm reportedC3Control level 3, Oversight evasionInternal researchDeveloper confirmed
Signal · Nov. 5
Capability2 sources
Signal · Nov. 16
Capability1 source
Signal · Nov. 16
Safety practice2 sources
Signal · Nov. 13
Policy2 sources
Signals · context, never counted
More on the horizon
All signals
Signal · Nov. 12
Frontier release
Signal · Nov. 12
Safety practice
Signal · Nov. 10
Capability
Signal · Nov. 10
Capability
Documented harms by report month. Selected reports, not a census.
New to Paperclip Index?
What we track, how the number works, and why it's named after a paperclip.
Start here