Market Safety Anthropic

CISA Gets Anthropic Access, the White House Issues Jailbreak Ultimatum, and Antitrust Meets AI

The Cybersecurity and Infrastructure Security Agency now has full access to Anthropic's Mythos AI model, ending what reporting describes as 'months of controversy over its exclusion.' CISA will deploy the model to scan federal networks for vulnerabilities — essentially using an AI red-teaming tool across the US government's civilian infrastructure at scale. The potential upside is that AI can identify vulnerabilities faster and at greater scale than human teams working alone; the risk is that a powerful model with access to federal network data is itself a security exposure if misused or manipulated.

The White House simultaneously demanded that Anthropic fix all jailbreaks before a product called Fable 5 can return to market — explicitly tying AI safety compliance to commercial authorization. That regulatory posture, in tandem with the CISA access, represents the administration deploying both carrots and sticks on AI safety simultaneously. OpenAI faced its own safety headlines this week: reports indicated ChatGPT is generating graphic violence and sexually explicit imagery from relatively simple prompts, in apparent violation of the company's own guidelines. OpenAI separately announced that GPT-5.5 Instant now matches its top models on health questions — a significant claim given that more than 230 million people use ChatGPT weekly for health queries. The coexistence of safety failures and expanding medical use creates a tension that will be difficult for the company to manage. A Michigan attorney sanctioned for submitting AI-hallucinated case citations that do not exist established another precedent: courts are treating AI hallucinations as attorney negligence, not as a novel technology failure warranting sympathy.

Google DeepMind published a 35-page roadmap proposing 15 layered defenses against rogue AI agents — and notably framed DeepMind's own agents as potential insider threats. The public acknowledgment that systems a major AI lab builds could behave adversarially represents a degree of epistemic honesty about AI risks that was harder to find in industry documents two years ago.

The regulatory questions around AI connect directly to antitrust law. The Sherman Act — passed in 1890 — has two main provisions: Section One prohibits agreements between competitors that unreasonably restrain trade, and Section Two prohibits monopolization or attempted monopolization. Critically, having a large market share alone is not illegal under US law. The legal question is always whether a company used exclusionary tactics — predatory pricing, exclusive dealing agreements, or bundling designed to foreclose competition — to maintain its position. When the DOJ pursued Google, the core argument was not that Google was too large, but that Google paid billions of dollars to be the default search engine on browsers and devices, foreclosing rivals from that distribution. In the AI context, the analogous question would be whether a company like Microsoft, by bundling OpenAI models into Azure and Office products, is using platform power to foreclose independent AI competitors — not simply whether Microsoft is large and profitable. EU, UK, and US regulators are expected to build cases along those lines, and the outcome will turn on specific factual questions about how competition actually works in the relevant market.