Nov 4, 2026

OpenAI Hires Cathedral as Apple Flags AI Agent Risks

OpenAI enlisted Cathedral, the startup that breached parts of Anthropic's systems, for a security sprint, while Apple warned investors AI agents could act unpredictably.

Listen to this update

The security of AI systems, and of the agents built on top of them, dominated the tech news cycle around Election Day. OpenAI brought in the startup whose researchers recently broke into parts of Anthropic's internal systems, Apple told investors that autonomous agents could misbehave in ways that are hard to undo, and fintech Revolut showed how far one company will go to rein in the cost of AI coding tools.

OpenAI recruits the team that breached Anthropic

OpenAI said on November 2 that researchers from Cathedral Systems will join its next internal security sprint. Cathedral's team will work alongside OpenAI staff and use OpenAI's own models to probe the company's defenses. OpenAI framed the move as a race against attackers, arguing that defenders need to put AI to work on vulnerability hunting "before attackers do."

The choice of partner raised eyebrows. Cathedral was founded by former Department of Government Efficiency (DOGE) staffers, with Gavin Kliger as CEO, and launched earlier this year at a reported $1.4 billion valuation after a $160 million round led by Andreessen Horowitz and Sequoia Capital. In late October, Cathedral told The Washington Post it had gained access to parts of Anthropic's internal environment, including Slack messages and meeting transcripts, during a bug-bounty engagement it said took under twelve hours.

Anthropic has pushed back on that account. Its security lead, Logan Graham, said the intrusion took several days, described Cathedral's methods as out of scope because they chained multiple third-party vulnerabilities, and said model weights and Claude conversations were never touched.

Alongside the partnership, OpenAI is sharply raising its bug bounty payouts. According to reporting on the announcement, rewards for critical bugs in core products now reach $30,000, while exceptional critical findings can earn up to $172,000 including bonuses. The move lands as OpenAI faces fresh scrutiny over internal security: a Reuters investigation published this week cited employees who say they warned for years about overly broad data access inside the company, claims OpenAI disputes.

Apple puts AI agent risk in writing

Apple used its latest annual filing to warn investors that AI agents could behave in "unintended or unpredictable" ways. The company said agents might exceed their permissions, resist instructions, or take actions that are difficult to reverse, potentially exposing Apple to legal and reputational harm. The filing does not say which agent technologies prompted the language or point to any specific incident.

The warning arrived the same day Apple opened its agent-style assistant to outside developers. A new Siri AI developer preview, shipping in the iOS 27.3, iPadOS 27.3, macOS 27.3 and visionOS 27.3 betas, lets apps expose content and actions through App Intents so users can ask Siri in plain language to check a delivery, book a table, or place an order. Developers can also build "skills" that chain several actions together. Apple says a public release is planned in a few months, with App Store submissions opening in the spring. The preview is not currently available in the EU on iOS, iPadOS or watchOS, and server-side Siri AI features carry daily usage limits.

Taken together, the two moves show Apple trying to have it both ways: expanding what its assistant can do inside other people's apps while formally flagging that autonomous software can go off script.

Revolut cuts AI coding bills by about 90%

On the cost side of the AI ledger, Revolut says it has cut inference spending on AI coding by roughly 90% in under a year. Speaking at the AI Engineer conference in New York, ops and infrastructure lead David Abitbol said the British fintech made self-hosted open-weights models its default instead of relying mainly on commercial tools such as Claude Code and OpenAI's Codex.

Revolut started with Z.ai's GLM-5.1 and has since moved to GLM-5.3, which it runs on 56 Nvidia B200 GPUs. According to The Register, self-hosting worked out to about a tenth of the cost of the commercial options, and rollout to roughly 2,000 engineers brought a reported 60% improvement in coding time. Abitbol said the biggest gains came not from writing code but from debugging and ticket work, because engineers no longer had to ration tokens. The approach has limits: GLM is text-only, so vision tasks are routed to other providers, and Abitbol called frontier models "the exception" rather than the rule.

Why it matters

As AI agents gain the power to act on their own, the industry is converging on the same uncomfortable lesson: the tools that make systems more capable also make them easier to attack and harder to control. OpenAI is betting that offensive AI expertise can harden its defenses, Apple is putting agent risk in front of shareholders, and companies like Revolut are showing that cost pressure is pushing enterprises toward models they can run and govern themselves.