Anthropic let an outside watchdog badge into the building
Issue #061 · 4 min read
Anthropic let an outside watchdog badge into the building
Plus: GitHub's autofix goes GA, GitLab locks down secrets, and Claude gets bored on a desert island
The big story
Anthropic just gave an outside watchdog the kind of access most companies won't even give their own board.
The group can publish what it finds about Anthropic's safety practices without asking Anthropic's permission first — that's Ars Technica.
Most companies that submit to outside audits keep a veto. They get to see the report before the world does, soften a line, delay a release, kill it if it's bad enough.
Anthropic just gave that veto away.
It's worth sitting with how unusual that is. A company that hands over "employee-like" access and then can't stop what comes out of it is a company betting that looking good under real scrutiny beats controlling the story.
Separately, Anthropic also said it's committing 5% of its shares to a new "public benefit fund" — that's TechCrunch.
Neither move tells you whether Anthropic's models are safer than anyone else's. Both tell you Anthropic has decided that "trust us" stopped working as a pitch, to customers, to regulators, to the enterprises deciding which lab to build on.
If you're picking a model vendor for anything regulated, the question worth asking isn't "what does your safety page say." It's "who can check, and what happens if they don't like what they find."
What shipped
GitHub's agentic autofix for code scanning alerts is now generally available. It reads through your codebase the way a developer would, finds the file that needs fixing, and proposes the patch — full details here.
Paired with GitHub's new AI Scan for pull requests, also GA this week and built to catch vulnerabilities in languages CodeQL doesn't cover, this is GitHub quietly building a full loop: scan, flag, fix, re-scan — see the AI Scan announcement. If your security backlog is the thing nobody has time for, this is the release to pilot.
GitLab's Secrets Manager just went GA, and the problem it solves is one every PM has lived through without naming it. A credential starts in an engineer's laptop, gets copied into a CI/CD variable to keep a pipeline running, and sits there as a raw value instead of a scoped reference — meaning a single signing key can reach every job in the pipeline, compromised or not. Here's how GitLab is closing that gap.
It shipped alongside a GitLab MCP server, now GA with GitLab 19.5, meant to stop agents from guessing at API endpoints and burning their context window — details on the MCP server. Worth noting: this is the third GitLab release this week touching how agents review and merge code, alongside a reviewer-matching tool and a governed code review flow. If your team runs agentic coding at any scale, GitLab is clearly betting the bottleneck isn't writing code anymore — it's reviewing it.
Vercel published its AI Gateway numbers for October, and one figure stands out: open-weight models now account for 71.5% of all tokens routed through the gateway, but only 23% of the spend — see the full index. That gap is the story. If you're still budgeting for frontier-model pricing across your whole roadmap, a lot of your usage may already belong on a cheaper model and nobody's told finance yet.
What I'd actually do this week
Ask your model vendor who audits them, and what that auditor is allowed to publish. Anthropic just set a new bar. See where your vendor sits against it.
Pull your last quarter's model spend and split it by open-weight versus frontier. Vercel's number suggests most teams are overpaying for tokens that don't need a frontier model. Check if yours is one of them.
If you're running agentic coding tools, look at GitHub's autofix GA and GitLab's secrets manager together. Fixing code and locking down the credentials around it are now both one-click GA features — pick one to pilot before your next security review.
Reply and tell me what you're finding. I read all of it.