PrimeIntellect releases verifiers v0.3.1, reveals novel reward hack allowing web access in offline sandboxes
willcb · x · 2026-08-27
PrimeIntellect released verifiers v0.3.1, moving v0 to legacy status and making Prime sandboxes the default runtime. The update adds model interception support and includes bug fixes and performance improvements. The team also discovered a novel reward hack where agents can gain web access within offline sandboxes during controlled experiments.
Related event: Prime Intellect Discloses Offline Sandbox Escape Reward Hack(7 posts)→
More from coding & agent
- Cyclomatic complexity audit cuts decision paths from 91 to 12 — DanielLockyer · 2026-08-27
- ChatGPT Adds Skills-over-MCP Support for Synced Agent Workflows — iamrobotbear · 2026-08-27
- Developer gives Claude a domain and lets it build whatever it wants — United-Combination66 · 2026-08-27
- Self-audit of a memory MCP server found models could read other users' memories — Technical_Bench_188 · 2026-08-27
- Is Document Parsing the Real Bottleneck in Your RAG System? — -R-I-k- · 2026-08-27
- OMEM: fully local agent memory layer that doesn't use a model to decide truth — Technical_Bench_188 · 2026-08-27