Agentic Misalignment

Framework / harness (not an eval — tooling/reference entry)

Scenarios where agents blackmail or leak under goal conflict or replacement threat.

Category

Deception, scheming & alignment auditing

Insurance exposure

Agent actions

Publisher

Anthropic

Publisher type

Frontier lab

License

MIT

Status

Stale (>12mo)

Last commit

2025-06-19

Stars

661

Notes

Basis of the 2025 'blackmail' findings.

View source on GitHub →

Wanich Risk Lab, 2026.