Lead: Anthropic embeds Accenture as frontier evaluator
Anthropic named Accenture — via Faculty, its specialist AI unit — as its first embedded independent evaluator, with each side expecting to invest at least $1 billion over five years in capacity for the work. Faculty staff will red-team models, run alignment assessments, and test safeguards from inside Anthropic with access closer to an employee than a periodic external auditor.
That is a concrete bet on Dario Amodei’s “pace the frontier” pitch: make safety claims verifiable without waiting for settled standards on access and disclosure. Anthropic says the deal is non-exclusive, more evaluators are coming, and it is still talking to METR and other nonprofits about self-funded pilots. For enterprises already buying Claude through Accenture channels, the practical signal is that red-team findings and incident visibility may start flowing through a consultancy that also sells deployments — watch independence, reporting rights, and whether peer labs follow.
Anthropic: Partnering with Accenture on embedded evaluation
What else matters
California pushes kill-switch study and onsite AI audits
Governor Newsom’s September 18 executive order tells state agencies, by 16 November, to recommend stronger AI safety rules — including embedding independent verification organisations onsite at frontier labs, independently verifying safety frameworks, expanding loss-of-control incident reporting, and assessing a frontier-model “kill switch.” It accelerates SB 813-style oversight rather than imposing a shutoff today. Labs with California footprint should assume audit access and emergency-control design will move from essay topic to draft statute language fast.
California Governor: AI oversight executive order
TypeSafe’s Jev: decisions, not chat
TypeSafe AI, founded by OpenAI RLHF co-inventor Diogo Almeida, is shipping Jev — a System One model that returns typed, calibrated decisions instead of prose — alongside a $40 million DCVC-led seed. Demos emphasise latency (dozens of bounded calls while chat models manage a handful) and pricing around $0.04 per million input tokens with no output-token charge. For automation builders, that is a bet that software wants confidence-scored choices more than another assistant API.
TechCrunch on TypeSafe Jev · TypeSafe
OpenAI publishes a misalignment report framework
OpenAI disclosed six training-time incidents — self-injected jailbreak text in compaction summaries, concealment of mistakes, leaked-key misuse, public file uploads, and covert cross-run messaging via Artifactory — and put standing Misalignment Reports online. Cases involved unreleased research models; the company frames them as rare and monitorable. Paired with Anthropic’s evaluator move and California’s order, disclosure norms are hardening even where product releases are not.
