RepoJournal
Safety Evals Safety Evals
52 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 2026-11-20
stories 49

© 2026 RepoJournal Home Showcase How it works Privacy

$ the-wire · showcase

enqueue_task resolves once per creator, Gemini 4 Neon support

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

Inspect fixes two enqueue_task() behaviors that were silently multiplying or hollowing out tasks added mid-run, and adds the model-info entry needed to drive gemini-4-neon's thinking controls correctly.

fix!: a task added with enqueue_task() from inside a running task runs once, under that task's model UKGovernmentBEIS/inspect_ai

by ransomr

A task added with enqueue_task() from inside a running task now resolves exactly once, under that task's model, instead of once per run model. Under eval() with several models that used to create N copies per creator, most of them on a model nobody asked for; this is a breaking change to that behavior.

fix: tasks added with enqueue_task() in eval-set selection workers run the real task, not an empty placeholder UKGovernmentBEIS/inspect_ai

by ransomr

In eval-set runs, selection-worker pruning used to stay active for the whole process, so a @task factory called by a running solver, tool, scorer or TaskSource returned an empty placeholder Task and enqueue_task() added that shell. The real factory now runs at the call, so task_with() changes survive.

feat(google): support Gemini 4 Neon UKGovernmentBEIS/inspect_ai

by brandly

Inspect gains a model-info entry for Gemini 4 Neon (gemini-4-neon), GA on the Gemini API as of 2026-11-18, whose thinking controls follow Hydrogen rather than Argon or Carbon. Two of them were being mapped wrong before this: xhigh and max were both sent as HIGH.

docs(design): sandbox agent bridge architecture and trust boundary, as built (#6619) UKGovernmentBEIS/inspect_ai

by Ransom Richardson

New as-built design docs for sandbox_agent_bridge() spell out the trust boundary: what the host does with a bridged request (model routing, generation settings, capabilities withheld by default), how approval and execution grants work, and what Inspect records.

docs(design): accepted residual risks of restic checkpoint egress (#6618) UKGovernmentBEIS/inspect_ai

by Ransom Richardson

The checkpoint snapshot design now states the accepted residual risks of restic checkpoint egress next to what the egress checks do not establish, including two ways a sandbox can make its own checkpoints fail to capture or restore.

Quick answers

What shipped in Safety Evals on November 20, 2026?
Inspect fixes two enqueue_task() behaviors that were silently multiplying or hollowing out tasks added mid-run, and adds the model-info entry needed to drive gemini-4-neon's thinking controls correctly. In total, 24 commits and 25 pull requests landed.
Who contributed to Safety Evals on November 20, 2026?
2 developers shipped this update, including Ransom Richardson and brandly.
What were the notable Safety Evals updates?
fix!: a task added with enqueue_task() from inside a running task runs once, under that task's model, fix: tasks added with enqueue_task() in eval-set selection workers run the real task, not an empty placeholder, and feat(google): support Gemini 4 Neon.