An OpenAI resignation reopens the debate over releasing increasingly powerful models

GOVERNANCE · UNITED STATES · reported 3–4 October

Reuters/WSJ: Jay Clayton to lead a new advisory AI panel

Reuters, citing the Wall Street Journal and an interview with Jay Clayton, reports that President Donald Trump has appointed the Director of National Intelligence as the administration’s new AI adviser. Clayton will reportedly lead a “Super Intelligence Force” with 120 days to produce a report on AI risks and opportunities.

The panel will review reporting mechanisms for breaches, hacks and other incidents and recommend ways to strengthen the federal response under existing authorities. This is an assessment and advisory mandate, not evidence that a new regulator with its own enforcement powers has been created. The details here are based on Reuters’ account of the WSJ report, rather than a separately verified White House order.

Why it matters AI is becoming more closely connected to national security and incident response. The important questions are what the panel recommends and what the government actually does with those recommendations.

AGENTS · PERMISSIONS · ANTHROPIC · 3 October; earlier fixes 2 October

Claude Code patches failures that could weaken an agent’s access boundaries

Claude Code 2.1.289 fixes several cases in which actions that should have been blocked or required approval were not handled as expected under particular configurations. These are changes to the software that runs the agent and controls its permissions, not improvements to the model itself.

One example involves compound terminal commands on managed machines with user-installed extensions. Another concerns read restrictions that were not enforced for some files accessed through symbolic links from the code editor. Anthropic also fixed a case in which a user-installed extension could change descriptions of sign-in tools from an organisation-managed MCP server.

The previous day brought fixes for certain dangerous deletion commands and checks that could be skipped when errors occurred. The releases follow the introduction of Claude Mods on 1 October, allowing extensions to change more aspects of Claude Code’s behaviour. The release notes document fixes, not confirmed exploitation by attackers.

Why it matters A model trying to follow instructions is not enough. The surrounding software must enforce the boundaries too. A rule saying that a file must not be read needs to hold regardless of the route the agent uses to reach it.

PRIVACY · DEVICES · APPLE · 2 October

Apple announces tighter controls for Full Disk Access

Apple has announced plans for additional controls around Full Disk Access in macOS. The permission gives an application broad access to protected data, including for purposes such as backup. It is not the same as unrestricted control over every function of the computer.

Apple says some developers use the capability in ways that may expose files, mail, messages and browsing history without users understanding its scope. The company therefore plans to require more explicit user action. Its announcement does not yet specify a release date or macOS version for the new controls.

The announcement follows public complaints about Meta’s Muse agent. Some users alleged it had read messages without the consent they believed they had given. Meta disputes this and says users must enable both Full Disk Access and the Messages connector. Apple has not concluded that Muse bypassed macOS permissions.

Why it matters When an agent can read documents, mail and messages and then act on that information, users need to understand what access means. Clearer approval matters, as does the ability to restrict and revoke permissions.

BUSINESS · SKILLS · 2 October

Anthropic commits $100 million to train 10,000 Claude deployment engineers

Anthropic has launched Claude Frontier Academy with a $100 million commitment and a target of training 10,000 Frontier Deployed Engineers by the end of 2027. These are an investment commitment and a training target, not money already spent or engineers who have already graduated.

The initial cohorts include employees from Accenture, Bain, Capgemini, Deloitte, McKinsey, Morgan Stanley, Novo Nordisk and Commonwealth Bank of Australia. The programme combines practical training with a twelve-week residency in which each participant leads a real Claude project in their organisation.

Anthropic describes a shortage of implementation expertise as a major bottleneck. The initiative can also be read as an ecosystem strategy: the vendor is training the people who will design, secure and operate systems around its platform.

Why it matters Organisations need more than access to a capable model. They need people who can integrate it and take responsibility for its operation. Vendor training can help, but may also deepen dependence on the same platform.