🧭 Anthropic Embeds Accenture as Its First Independent AI Evaluator — $1 Billion Each Over Five Years

Anthropic has announced a partnership with Accenture — specifically its specialist AI unit Faculty — to become the first embedded evaluator inside Anthropic. The arrangement puts independent evaluators inside the company with "access comparable to an employee's", allowing them to observe model development during training, monitor deployment decisions in real-time, identify blind spots in Anthropic's operations, and file public incident reports. Anthropic and Accenture each plan to invest at least $1 billion over five years to build capacity in this area; Anthropic will directly fund Accenture's work at the company. The partnership is non-exclusive: Anthropic is simultaneously engaging METR and additional evaluators to be announced in coming weeks, and Accenture will offer the same embedded-evaluation service to other AI labs.

What embedded evaluators will do

• Red-team models before release: Faculty evaluators will access unreleased model weights and run adversarial probes across safety, capability, and alignment dimensions — with full visibility into training runs, not just end-product APIs.
• Conduct alignment assessments: Structured reviews of whether model behaviour matches stated values; evaluators can flag findings directly to Anthropic's board under the arrangement, not just to the safety team.
• Test safeguards in deployment: Ongoing monitoring of production model behaviour, with escalation authority to pause rollouts if critical issues are discovered.
• Report incidents publicly: Evaluators are explicitly empowered to publish findings — this is the key structural difference from an internal red team, whose results can be withheld.

🧭 Claude Code v2.1.277 Adopts AGENTS.md — AI Coding Agents Converge on a Shared Instruction File Standard

Claude Code version 2.1.277 (released September 18) adds native AGENTS.md support: if a project directory contains no CLAUDE.md, Claude Code now checks for and reads an AGENTS.md file instead. AGENTS.md is a community-driven instruction format popularised by OpenAI Codex that has been adopted across more than 60,000 public repositories. By adding first-class support for it, Anthropic is making a deliberate move toward an ecosystem where different coding AI agents can collaborate on the same codebase without requiring teams to maintain duplicate or agent-specific instruction files. The behaviour is a conservative fallback — repos that already have a CLAUDE.md are unaffected — and it is configurable under /config → Project instructions.

How to use it and what to expect

• Zero migration required for new repos: Create a single AGENTS.md and all supporting coding agents (Claude, Codex, Gemini CLI etc.) will read it — no agent-specific files needed.
• Existing CLAUDE.md repos are safe: The fallback only triggers when CLAUDE.md is absent; there is no merging or override of your existing instructions.
• Configurable behaviour: Open /config and look for Project instructions to toggle or override the AGENTS.md fallback on a per-project basis.
• Not yet on Bedrock, Vertex, or Foundry: The feature requires Claude Code's local MCP infrastructure; managed cloud deployments will follow in a later release.

# AGENTS.md — shared across Claude Code, Codex, Gemini CLI
# Placed at repo root; any participating agent reads it

## Project context
This repo implements a FastAPI service. Target Python 3.12.
All code must pass `ruff check --fix` before committing.

## Agent conventions
- Never commit directly to main; always open a PR
- Tests live in tests/; run `pytest -x` before proposing changes
- Use existing helper functions in utils/ before writing new ones

## Out of scope
Do not modify pyproject.toml or Dockerfile without explicit instruction.

🧭 Compliance API Gains Full Cross-Surface Coverage: Claude in Chrome, Claude Science, and Microsoft 365 Now in Beta

Anthropic has expanded the Compliance API's local session endpoints to cover three additional Claude surfaces that were previously invisible to enterprise compliance teams: Claude in Chrome (the browser extension), Claude Science (the research-tools surface), and Claude for Microsoft 365 (Excel, PowerPoint, Word, and Outlook). The expansion is in beta for Claude Enterprise organisations; no new credentials are required — it uses the existing Compliance Access Key and read:compliance_user_data scope that teams already use for Claude.ai and API transcript retrieval.

Querying the new surfaces

• List sessions across all surfaces: GET /v1/compliance/apps/sessions/local — returns sessions from Claude.ai, Claude in Chrome, Claude Science, and MS365 in a single paginated response.
• Retrieve one session's metadata: GET /v1/compliance/apps/sessions/local/{session_id} — includes a product_surface field so you can filter by surface downstream.
• Fetch the full transcript: GET /v1/compliance/apps/sessions/local/{session_id}/messages — returns the same transcript format as existing endpoints; parse product_surface: "claude_in_chrome", "claude_science", or "ms365" to distinguish.

import anthropic

client = anthropic.Anthropic(api_key="...", compliance_key="CAK-...")

# List sessions from the last 7 days (all surfaces)
sessions = client.compliance.apps.sessions.local.list(
    created_after="2026-09-14T00:00:00Z"
)

for s in sessions.data:
    if s.product_surface in ("claude_in_chrome", "claude_science", "ms365"):
        transcript = client.compliance.apps.sessions.local.messages.list(
            session_id=s.id
        )
        # Process transcript for your SIEM / DLP pipeline