Finding and fixing software bugs

  • A PostgreSQL maintainer says AI-found bugs are slowing the next release — Finding defects faster can move the bottleneck to human review and repair.
    • Major contributor Robert Haas expects PostgreSQL 19 to arrive in November or later. Three major features were removed late in the beta period.
    • Between April 8 and September 25, commits to the release branch rose from 724 in 2025 to 1,169 in 2026. Fixes applied to the prior stable branch rose from 248 to 433.
    • Haas says people are reporting many bugs found with language models. Generated fixes still need human review, and security and older-branch fixes compete with release work.
    • Caveat: Haas offers a maintainer’s explanation for the delay; the commit counts do not independently establish how many fixes originated with AI. (HN 467 points · 295 comments · maintainer’s account)

Safer agent workflows

  • A recovery study tests what to do when a tool times out paper notes → — A timed-out action may still have happened, so checking its effects can prevent duplicate work.

    • An agent is a model running a multi-step task with tools. The paper considers checking remote state, waiting, retrying with a valid deduplication key, compensating a change or handing work to a person.
    • In 37 modeled assignments, planning and three simpler active policies each certified 28 safe completions. The authors’ planner saved four synthetic fee units while using more computation.
    • In a separate test of two local model bundles, planning advice produced 30 certified completions versus nine with evidence prompting across 58 paired cases. Both conditions used the same action-safety check.
    • Caveat: The planner receives a supplied finite model of possible provider behavior; the results do not guarantee exactly-once execution against arbitrary services. (arXiv 2610.41453 · full paper)

  • Search tools can find changes that assistants never think to check paper notes → — Ask an assistant to verify time-sensitive assumptions, even when its answer includes search results.

    • Researchers tested eight models with direct questions about changes and indirect requests that required noticing the same changes. A benchmark is a fixed set of tasks for comparing models.
    • For deaths of public figures, indirect accuracy fell by a median 27.5 percentage points at high reasoning effort in the models’ own search setups. Direct questions reached at least 96.5% accuracy in nearly all configurations.
    • The released code and dataset support the test. The study also examines changes in public officials’ tenures.
    • Caveat: The measured dataset covers deaths and ended tenures; the paper’s museum-closure example illustrates the idea but is not a measured category. (arXiv 2610.41435 · full paper)

Testing models

  • SemiAnalysis launches a broad test of models doing specialist work — The results are a useful reminder to inspect task design and tool setup before treating a model ranking as portable.
    • GDPx lists 206 tasks across eleven domains, including software engineering, chip design and security, supplied by fourteen specialist partners.
    • The authors report a top overall score of 12.0 for Claude Opus 5.5. Domains receive equal weight, with three attempts per task used for the headline score.
    • Most models used their own tool-running software when available; others used OpenCode. The default task time limit was eight hours, with shorter limits for some tasks.
    • Caveat: The ranking compares complete model-and-tool setups on the authors’ chosen tasks, and the partners include companies selling evaluation environments. (Techmeme · GDPx results and methods)

Deepfake abuse

  • Researchers find abusive deepfake tools hidden behind ordinary-looking apps — The investigation shows how app-review checks can miss harmful features delivered after approval.
    • WIRED’s investigation traces more than 400 connected domains and apps, some presenting as pet-portrait or other ordinary services.
    • Researchers found the apps could load an explicit interface under certain conditions, including apparently when installed after an associated advertisement. Users could upload photos to generate abusive images.
    • WIRED reports that Apple removed 19 apps after being contacted. Google had removed most of the identified apps before the inquiry; the researchers reported over 350,000 collective downloads for 27 apps live on Google Play in September.
    • Caveat: These are researchers’ and WIRED’s findings about one linked network, not a count of all abusive apps across both stores. (Techmeme · AI Forensics investigation)
5 gathered items not included