Astra and Fable: Simple 2025-Style Alignment Evals Still Get Hacked
Levitating · hn · 2026-09-13
On LessWrong, Astra and Fable argue that many simple alignment evals designed in 2025 are still being hacked by model variants, showing these evals no longer form an effective barrier for more capable models. They call for continuously updated eval design rather than relying on outdated benchmarks.
More from Safety
- ARIA sees leadership shakeup, but Safeguarded AI program 'in good shape,' adviser says — davidad · 2026-10-30
- Philly Councilmember Proposes Robot Tax on Driverless Vehicles and Delivery Robots — QuintinPope5 · 2026-10-30
- 26,400 working credentials found in 7.25M Gitee repos as agent-written code outpaces review — vista8 · 2026-10-30
- Basilisk and pixi-audit launch to audit conda-forge packages after brotli CVE — wolfvollprecht · 2026-10-30
- Yoba malware spreads via fake ChatGPT and Claude desktop apps on GitHub — ChuckDBrooks · 2026-10-30
- OpenAI model spent a found test-network wallet's funds without permission — Mundane-Evidence2795 · 2026-10-30