Security researchers: nation-states already deploying abliterated open-weight models for offense
joshua_saxe · x · 2026-10-22
- Security researcher kuza55 pushes back on the assumption that nation-states lag on AI adoption for offensive use, noting offensive capability work has long been outsourced to private industry.
- Nation-states lack access to frontier US models but are already adopting abliterated (safety-removed) GLM-5.3 and similar open-weight models, and have better access to talent.
- Joshua Saxe largely agrees but argues nation-states will still trail companies like Google on the adoption curve.
Related event: Researchers Debate Whether State Actors Already Use Uncensored Open Models(3 posts)→
More from Safety
- Debate: Can Anyone Independently Verify METR's AI Evals? Critics Say No — MoonL88537 · 2026-10-23
- php2shell-poc open-sources two CVSS 9.2 pre-auth PHP RCE exploits affecting 14 years of releases — evilsocket · 2026-10-23
- Security Through Binary Patching: fixing critical PHP CVEs with radare2 in 24 bytes, no recompile — SIGKITTEN · 2026-10-23
- Geoffrey Irving: labs aren't playing 4D chess — do simple, correct things so others update — geoffreyirving · 2026-10-23
- Anthropic researcher's human-written essay flagged 92% AI by detector Pangram, sparking drama — sethlazar · 2026-10-23
- 86.7% of exploited CVEs are now zero-days: Kurtz on why PHP2Shell ends the patching window — joshua_saxe · 2026-10-23