RepoJournal
Safety Evals Safety Evals
24 wires and counting

$ follow Safety Evals

Keep up with Safety Evals in about 3 minutes: what actually shipped — the commits, pull requests, releases, and security advisories that matter.

or

fair warning: these emails are deeply technical. diffs, version numbers, CVEs, benchmark deltas. if that's not your idea of a good read, this isn't your newsletter.

Folds into your digest — weekly by default, monthly if you prefer. Unsubscribe in one click.

$ status

wire 2026-10-24
stories 61

© 2026 RepoJournal Home Showcase How it works Privacy

$ the-wire · showcase

inspect_evals 0.26.0 drops GAIA's browser tools; background subagent failures now fail samples

By RepoJournal · Filed · About Safety Evals · Composed from the cited sources · methodology

inspect_evals v0.26.0 changes what the default GAIA agent can do, while inspect_ai fixes three failures that could quietly corrupt results: swallowed background subagent errors, a broken Windows file:// path, and host-timezone drift in analysis timestamps.

v0.26.0 UKGovernmentBEIS/inspect_evals

by github-actions[bot]

The default GAIA agent no longer has deprecated browser tools; it now uses web search, with `web_search_providers` controlling provider selection. Defaults to built-in model search, with Tavily as fallback requiring `TAVILY_API_KEY`, and Anthropic search needs Inspect 0.3.278 or later. If your GAIA runs relied on interactive browser actions, they are gone and results are not comparable across t...

fix: fail samples on unexpected background subagent errors UKGovernmentBEIS/inspect_ai

by XCODESSS

Unexpected exceptions raised by background subagents used to be absorbed, so a sample kept scoring after an infrastructure failure that would have killed a foreground subagent, for instance an agent working against an empty filesystem after its sandbox was replaced. Background dispatch now propagates those failures to the sample runner, so `fail_on_error` and `retry_on_error` apply to them as t...

fix: local_path no longer drops the Windows drive from file:// URIs UKGovernmentBEIS/inspect_ai

by Cross2pro

On Windows, `list_eval_logs()` and `resolve_logs()` returned `file://D:/logs/...` URIs whose drive letter landed in the URL netloc, so `local_path()` dropped it and resolved against the current drive. `read_eval_log`, `read_eval_log_headers`, `evals_df`, and `samples_df` raised on those logs as a result; they now carry the drive through.

fix: keep timezone-less dataframe timestamps on the UTC clock (#5767) UKGovernmentBEIS/inspect_ai

by Apurv Gupta

Timezone-less timestamps read through dataframes are now treated as UTC rather than being rescheduled onto the host clock, so analysis results no longer shift by the operator's local offset. The regression test runs on a +05:30 host, which is the case that used to pass silently on UTC CI.

ci: pin the release build and npm toolchains UKGovernmentBEIS/inspect_ai

by dragonstyle

The remaining work is supply-chain hygiene: the release build and npm publish steps pinned their toolchains, so an unpinned `build`, `setuptools`, or npm tool could no longer alter published artifacts. Nothing changes for operators, but it is what stands between you and a tampered release.

Quick answers

What shipped in Safety Evals on October 24, 2026?
inspect_evals v0.26.0 changes what the default GAIA agent can do, while inspect_ai fixes three failures that could quietly corrupt results: swallowed background subagent errors, a broken Windows file:// path, and host-timezone drift in analysis timestamps. In total, 29 commits, 31 pull requests, and 1 releases landed.
Who contributed to Safety Evals on October 24, 2026?
7 developers shipped this update, including XCODESSS, Cross2pro, Apurv Gupta, dragonstyle, rasmusfaber, github-actions[bot], and github-actions.
What were the notable Safety Evals updates?
v0.26.0, fix: fail samples on unexpected background subagent errors, and fix: local_path no longer drops the Windows drive from file:// URIs.