$ cat briefs/daily/2026-11-01.md
2026-11-01
November 1, 2026 (Sun)
1 security report · 2 paper picks · 1 creative workflow
Top Stories
1. A patched ChatGPT flaw crossed the boundary between features
Parameter Labs reports that a public share could expose private conversations through interaction between HTML Preview and a privileged graph component. Its October 28 disclosure says OpenAI confirmed the fix on 2026-10-26. This was a controlled demonstration, not a report of observed attacks against users.
Why it matters: isolating generated code is insufficient if another component trusts messages from it. → Agent Runtime Containment (source)
Paper Picks
WorkFactory: probe and repair professional-task environments before collecting demonstrations. Authors construct 112,686 tasks across 612 occupations; WorkPilot-35B-A3B rises from 879 to 1565 Elo on GDPval-AA v2.1. Read it for executable task design; total construction cost is unspecified in the abstract. → WorkFactory — executable tasks before expert demonstrations (source)
A-RSI: refresh training tasks when learning signals weaken, using program-derived answers instead of majority labels. The reported suite average rises from 57.77 to 61.05 with 800 fresh tasks. Read it for separating refresh decisions from answer checks; the abstract leaves audit coverage and total compute unclear. → A-RSI — refreshing training tasks without using test scores (source)
Watch
New in Wiki
- WorkFactory — executable tasks before expert demonstrations and A-RSI — refreshing training tasks without using test scores add methods, authors and evidence limits. (tasks) (refresh)
Updates
- Eval Harness Configuration records Arena's 2026-10-28 publisher date: Claude Opus 5.5 (Max) has 20.94% Net Improvement, 95% CI 18.79% to 23.09%. This is neither Elo nor task-success probability. (source)