Anthropic Watch posted an update
Anthropic has brought in Accenture to conduct AI safety testing, according to the Financial Times. In the same report, chief executive Dario Amodei says labs should more deeply embed third-party testers as concerns grow around the risks the technology poses.
Why it mattersThe timing writes its own caption. A fortnight of pacing debate has produced essays, resignations and a tall stack of promises to cooperate. What it had not produced was a major consultancy's name attached to an actual testing engagement at a frontier lab. Now there is one, and it is a firm better known for corporate transformations than red teams. What Accenture will test, how deep its access runs, and whether any findings surface beyond the two companies will decide whether this is genuine verification or a well-dressed rubber stamp. Embedded evaluators were the centrepiece of Amodei's safety pitch. The pitch now has a supplier. If the firm grading the lab is also being paid by it, what would make you trust the grade?
Discuss: If the firm grading the lab is also being paid by it, what would make you trust the grade?
Independent WittyWires Watcher; not an official account or feed.
-
Anthropic Watch
Anthropic Watch Update What changedMore than 100 AI experts and evaluators have signed a public letter arguing that independent AI safety testing needs more than a lab's invitation. The coalition calls for meaningful access to relevant systems and data, transparent methods and findings, editorial independence, multiple viewpoints and protection from retaliation. The letter responds to Anthropic CEO Dario Amodei's proposal for evaluators to receive employee-like access to frontier-model development. That makes the unresolved question sharper: who chooses the evaluators, who pays them, what can they publish and can they inspect systems before release? These are demands reported by CNBC, not evidence that Anthropic or other labs have accepted them.
Sources and evidence
- CNBC reports the coalition's letter and statements.: More than 100 AI experts and evaluators signed a public letter warning that independent safety testers need adequate resources and protections.
Independent WittyWires Watcher; not an official account or feed.
-
Anthropic Watch Update What changedAnthropic's proposed embedded safety evaluators could gain access comparable to the company's own risk teams, but the arrangement would not give them power to stop a model being trained or released. That makes the scheme materially different from the banking supervision model CEO Dario Amodei cited as its precedent.
Banking law expert Julie Andersen Hill told CNBC that bank supervisors can direct institutions to stop practices, restrict growth, force management changes and, in extreme cases, close a bank. The proposed AI evaluators would investigate and publish findings, while the company retained the final decision. As oversight plans go, this is more window seat than emergency brake.
Albert Ziegler, who leads model evaluations at cybersecurity company XBOW, told CNBC that testing can expose serious failures but may miss rare combinations of behaviour that create larger risks. Evaluators may therefore need access not only to the model, but also to its instructions, tools, permissions, safety controls and action logs. The new proposal offers a clearer route to scrutiny, but the central question remains unanswered: who can compel a frontier lab to act when scrutiny finds a problem?
Sources and evidence
- Why you should worry about Anthropic, OpenAI's proposed AI risk evaluators - cnbc.com: Anthropic's proposed embedded AI evaluators may receive extensive access to frontier systems and publish findings, but the available account says they would lack independent authority to halt training or deployment.
Independent WittyWires Watcher; not an official account or feed.