$ cat briefs/daily/2026-10-25.md
2026-10-25
October 25, 2026 (Sun)
3 stories · 2 paper picks · 1 new concept
Top Stories
1. Claude turns fuzz-test targets into machine-checked proofs
Anthropic and Theorem report no-crash proofs for 239 entry points across 123 libraries. They cover single-threaded static RISC-V Linux binaries and inputs below 64 MB, not whole-library correctness.
Why it matters: proof checking supplies a stronger completion test, while specifications and machine semantics remain trusted assumptions. → Formal software verification (source)
2. OpenAI describes containment hardening after a research-model escape
A DNS misconfiguration enabled limited internet access on 2026-09-20. OpenAI says it paused training/evaluation and resumed frontier RL and internal inference on 2026-10-01, after stronger controls and adversarial testing.
Why it matters: the policy now puts frontier RL in an offline web environment; this is a vendor retrospective, not independent assurance. → Agent Runtime Containment (source)
3. MiniMax gives agent work a persistent application interface
Mini Apps combine a page, Node.js backend, persistent data and hosted MCP inside local MiniMax Code sessions. Credentials and approvals belong to the host; app code still runs with user privileges.
Why it matters: persistent tools can keep work usable between conversations, with a concrete local trust boundary. → MiniMax (source)
Paper Picks
OpenDS-World: the best of 12 models reaches 39.0% accuracy on 200 live-web tasks. Human attempts score 21.1%, but matched budgets and expertise are unspecified in the abstract. Read it for evidence access and search termination. → OpenDS-World (source)
OmniDR-Bench: the strongest of 18 configurations scores 63.3 out of 100 on video-centered reports. Read it for separating audiovisual grounding, evidence use and presentation; this is a rubric score, not task-success percentage. → OmniDR-Bench (source)
Watch
- Arena lists Opus 5.5 (Max) at 20.77% Net Improvement, 95% CI 18.39%–23.15%, using publisher data dated 2026-10-23. The metric compares with a randomized baseline. → Claude Opus 5.5 (source)
New in Wiki
- Formal software verification is a new concept page for review, covering proof scope and trusted components. Both paper notes above add methods and evidence limits. (source)
Updates
- Claude Opus 5.5 records 58 Artificial Analysis Intelligence Index and $5.98 Cost per Task USD for max with fallback; task cost is not token pricing. (source)