$ cat the-wire/2026-10-23.log

The morning wire
on AI safety.

Every day, RepoJournal files an AI-written briefing on what shipped across the tools keeping models honest — Inspect, promptfoo, Guardrails, TransformerLens, and more. Today's wires:

ai-safety-evals

$ ls orgs/ --wires=today

Safety Evals

live

Inspect logs stop mutating mid-run, Math.Calc regrades

inspect_ai fixed a logging bug that silently rewrote the record of what every model saw, while inspect_evals changed how Math.Calc grades, so scores from the infinite_bench benchmark can move for reasons unrelated to your models.

wire 2026-10-23

shipped 23

Inspect, Petri, and Control Arena - how frontier labs and safety institutes evaluate and audit models Read →

AI Red Teaming

live

PyRIT calms Rust false positives and fixes attacker-view reporting

PyRIT spent the day repairing how its own results are reported: the default Rust scenario stops flagging standard-library paths as hallucinated packages, and scenario-result views stop surfacing orchestration envelopes to operators reviewing attack runs.

wire 2026-10-23

shipped 48

promptfoo, PyRIT, garak, and DeepTeam - adversarial testing for LLM apps and agents Read →

Agent Security

live

Windows push keys survive uninstall, and managed installs can skip absent agents

snyk/agent-scan shipped two operator-facing changes: a fix so guard uninstall actually revokes the push key on Windows, and an opt-in guard install --check-installed flag for managed (MDM) installs.

wire 2026-10-23

shipped 16

Guardrails, sandboxes, and scanners - securing AI agents at runtime Read →

Interpretability

live

TransformerLens 4.3.0 lands with fixes for J-space projections, ablation, and bridge block discovery

TransformerLens v4.3.0 shipped alongside fixes that correct two silently wrong analysis results and one silent empty-result path in the bridge, all of which affect numbers you may already have.

wire 2026-10-23

shipped 10

TransformerLens, SAELens, and nnsight - opening up what models are actually doing Read →

Alignment & Post-training

live

Open-Instruct moves to Miles v0.1.2; verl fixes teacher top-k packing under CP

The only two changes an operator can act on today are open-instruct's runtime move to Miles v0.1.2 and a verl fix for teacher logits that matched the student's shape but not its tokens; the rest is CI and device-abstraction plumbing.