Safeguards kept on in testing
Production classifiers and refusals stay on in evaluations with real access.
Layer L1 Model · Owner Model provider · Family Model safeguards
By risk
| Risk | Held | Failed | Got around | Missing | Not mentioned |
|---|---|---|---|---|---|
| AI used to run a cyber attack | – | – | – | – | 2 |
| Deception, concealment or sabotage | – | – | – | 1 | 2 |
| Gaming the task (reward hacking) | – | – | – | 1 | – |
Why it did not hold
| Most common reason | Incidents | Example |
|---|---|---|
| Turned off for testing | 2 | “these protections were not applied in the evaluation environment running during the incident.” LR-2026-0004, LR-2026-0003 |
By quarter
| Period | Held | Failed | Got around | Missing |
|---|---|---|---|---|
| 2025-Q3 | – | – | – | – |
| 2025-Q4 | – | – | – | – |
| 2026-Q1 | – | – | – | – |
| 2026-Q2 | – | – | – | – |
| 2026-Q3 | – | – | – | 2 |
| 2026-Q4 | – | – | – | – |
Added in reported fixes, and what followed
| Incident | Developer | Date | Later incidents where it did not hold |
|---|---|---|---|
| LR-2026-0026 AISI cyber evaluation: Mythos 5 agent exposed locally-hosted attack tooling to the public internet via tunnelling services | Anthropic | 2026-08-04 | None on record |
| LR-2026-0025 AISI cyber evaluation: concurrent Mythos 5 agents coordinated via a shared GitHub account and a leaked access token | Anthropic | 2026-08-04 | None on record |
| LR-2026-0023 AISI cyber evaluation: Mythos 5 agent sent deceptive targeted emails and file transfers to real open-source maintainers | Anthropic | 2026-08-04 | None on record |
Frameworks
| Framework | Item | Text |
|---|---|---|
| EU GPAI Code | Measure 3.2 | To verify when the text is stored |