Claude News
FR
Live

Claude, the essentials — edition of September 19, 2026

Anthropic Brings Outside Evaluators In-House While Loosening Rules for Science

By ·

Anthropic detailed a billion-dollar plan to embed independent evaluators inside the company and opened a verification program that eases safety limits for vetted life-science work, while Claude Code shipped four rapid releases hardening enterprise gateway controls.

Key points

Outside eyes move inside the lab

Anthropic announced a partnership with Accenture, led by Accenture's specialist AI unit Faculty, to build what it calls embedded evaluation: independent reviewers evaluating and red-teaming Anthropic's models, running alignment assessments and testing safeguards from a position inside the company rather than outside it. Anthropic frames this as a concrete step toward a commitment made in its CEO's essay "We Must Pace the Frontier," and says it and Accenture each expect to invest at least $1 billion over the next five years. The access described goes well beyond a typical external audit: Anthropic says embedded evaluators will have access "comparable to an employee's," letting them watch models take shape during training, follow the internal decisions that govern how models are built and deployed, and speak directly to Anthropic staff — access meant to let them assess how the company operates, verify it is keeping its safety commitments, spot blind spots, and report incidents.

Anthropic is explicit about the limits of this arrangement, a framing worth preserving rather than smoothing over: it states that independent embedded evaluators "do not reduce our accountability, but help to make it more verifiable," and that responsibility for model safety remains its own. The company also acknowledges the practice is immature — there are no standards yet for what access embedded evaluators should get or how they should report findings, and no settled system for funding independent evaluation. Anthropic's stated preference, laid out in its June Advanced AI Framework, is that such funding should eventually come from pooled or government sources; because neither exists today, it is funding Accenture's work itself while also in dialogue with METR and other nonprofit evaluators about piloting elements of the model using their own funding instead. A companion piece published the same day on measuring the pace of development inside frontier labs reflects the same broader push toward externally legible metrics of how fast and how safely AI companies are moving.

Sources: Partnering with Accenture on embedded evaluation — Anthropic · Measurements for understanding the pace of AI development inside frontier labs — Anthropic

A tailored gate for life-science work

Alongside the evaluation push, Anthropic introduced the Life Sciences Verification Program (LSVP), which gives vetted life-science professionals access to Mythos, Opus and Sonnet with safeguards recalibrated to permit biology-related work — drug discovery, research biology, clinical development, manufacturing — that remains blocked on Anthropic's generally available Fable models. The program launches in beta for teams and institutions, building on an early-access phase that has already onboarded dozens of organizations; applicants go through a verification process covering research credentials, security standards and ethical research oversight, and access extends across Claude Science, Claude.ai, Claude Code and the API.

The program is structured in two tiers rather than as a blanket loosening of restrictions. "Standard Use" grants, renewed annually, cover the bulk of biology R&D and adjacent activity — basic science, supply chain and manufacturing, clinical development, quality assurance, regulatory affairs, investing and diligence — and apply to Mythos 5.1, Opus 5 and Sonnet 5 today, extending automatically to future models. "High-risk Use" grants are reserved for work Anthropic judges carries greater potential for misuse and require additional vetting. Taken together with the Accenture partnership, the day's announcements point to the same underlying approach: calibrated, verifiable trust extended to vetted parties, rather than uniform rules applied to everyone.

Sources: Introducing the Life Sciences Verification Program — Anthropic

Claude Code's gateway plumbing, four releases in a row

Claude Code moved through four point releases in quick succession. Version 2.1.275 added a confirmation step showing the signed-in account when a Claude apps gateway names one during sign-in, before the credential is saved and visible in /status, alongside a send-now key (ctrl+enter, or ctrl+x ctrl+s) that interrupts the current turn to send queued messages immediately, and a startup warning when a configured otelHeadersHelper fails so telemetry doesn't silently stop exporting. Version 2.1.276 fixed a regression introduced in 2.1.275 that caused every request to fail with a 400 error referencing the input tag 'advisor_20260301' whenever ANTHROPIC_BASE_URL pointed at a proxy or gateway — the tag's name suggesting the bug surfaced through feature-flag plumbing tied to the same period as the Financial Advisors launch. Version 2.1.277 added support for the community AGENTS.md convention as a fallback when a project has no CLAUDE.md, plus two gateway options for enterprises running their own forward proxies: an egress-boundary flag that hands proxies a hostname instead of resolving it locally, and a headers map for sending static headers to a proxy in front of a provider. Version 2.1.278 changed auto mode's default for API, Enterprise, Bedrock, Vertex, Foundry and gateway users to a server-side classifier that Anthropic says is not charged for classifier overhead, with an opt-out variable and a new /status row showing whether a session's classifier is running server-side.

Read together, the run of releases shows Anthropic tightening the operational plumbing enterprises rely on to route Claude Code traffic through their own infrastructure — proxy egress, static headers, gateway credential confirmation — while also making a billing-relevant default change transparent in /status, and shipping a same-week fix for a regression its own prior release had caused.

Sources: Claude Code 2.1.278 release notes — Claude Code · Claude Code 2.1.277 release notes — Claude Code · Claude Code 2.1.276 release notes — Claude Code · Claude Code 2.1.275 release notes — Claude Code

Compliance catching up with new surfaces, and further vertical pushes

Anthropic's Compliance API was extended so that local session endpoints now also return transcripts of Claude in Chrome sessions, tagged with the product_surface value claude_in_chrome, in beta for Claude Enterprise organizations using their existing Compliance Access Key and the read:compliance_user_data scope. The update means compliance and audit tooling is being extended to cover Claude's browser-agent surface as it reaches enterprise customers, following the same pattern as the day's other announcements: new capability paired with the oversight infrastructure to govern it.

Anthropic also published material on Claude for Financial Advisors and on how Claude is being used in biomolecular modeling, and announced Claude Camp for Customers. Taken with the Life Sciences Verification Program, the day's announcements collectively point to continued specialization of Claude around regulated and technical verticals — finance, biology, and customer-facing enablement — alongside the compliance and evaluation infrastructure needed to support that expansion.

Sources: Inside Claude for Financial Advisors — Anthropic · Compliance API release notes — September 18, 2026 — Anthropic · Claude Camp for Customers — Anthropic · How Claude is uplifting biomolecular modeling — Anthropic

This edition is an original synthesis written by Claude from aggregated news — Anthropic's own sources first (release notes, status, newsroom, research, engineering), then the press, Hacker News, Reddit and GitHub, under the editorial supervision of Héra SASU. Every fact links to its article, publisher named. See the live feed →

Claude News is published by Héra SASU. Independent media, not affiliated with Anthropic.