VC
Value Add VC
⚡HomePulse⚡Helpful Apps📝Blog🤝Partner
Illustration for: The Fed Sounded the Alarm on Anthropic's Mythos -- Without Access to It
Value Add VC/Pulse/REGULATION

The Fed Sounded the Alarm on Anthropic's Mythos -- Without Access to It

New reporting reveals the Federal Reserve warned major bank CEOs about unprecedented cybersecurity risks from Anthropic's Mythos AI model months ago, while the central bank itself still lacked access to the tool as of mid-July.

By the Numbers

April 2026
Initial Fed/Treasury warning
June 2026
Export controls imposed
Jun-Jul 2026
Export controls lifted
Still none
Fed access status (Jul)
TC
Trace Cohen
Early-stage VC & angel · Founder, New York Venture Partners
July 21, 2026
2 min read
ShareXLinkedInEmail

THE RUNDOWN

1

In April, the Federal Reserve and Treasury Department convened an extraordinary meeting with the CEOs of the nation's top banks to warn about Mythos, an Anthropic model demonstrating unprecedented ability to find and exploit previously unknown zero-day vulnerabilities in software

2

Officials reportedly believe Mythos could debilitate Fortune 100 companies, infiltrate national defense systems, or take down large portions of the internet if misused -- concrete enough concern that it triggered a direct CEO-to-CEO warning process rather than a written advisory

3

In June, Anthropic disabled access to Mythos 5 and its broader release Fable 5 to comply with federal export-control directives citing national security, before Commerce Secretary Howard Lutnick restored access for a select group of 'trusted partners' and the controls were later lifted entirely

4

As of mid-July, new CNBC reporting shows the Federal Reserve itself still lacked access to Claude Mythos Preview -- meaning the regulator warning banks about the model's cyber risks couldn't independently test or verify those risks with the actual tool

TC

The VC Read · Trace's Take

Trace Cohen

A regulator sounding the loudest public alarm about a model's cyber risk while having zero hands-on access to test it is a genuinely bad look for how seriously the government's own AI-safety infrastructure is being resourced. If the Fed can't get access, most enterprises evaluating frontier-model risk can't either -- which is exactly the gap agent-security startups like Neo and Empirical are racing to fill with third-party tooling instead.

Analysis

New CNBC reporting reveals an awkward gap at the center of Washington's response to advanced AI cyber risk: the Federal Reserve warned the CEOs of the nation's top banks in April about unprecedented cybersecurity threats from Anthropic's Mythos model, while the Fed itself still lacked direct access to the tool as of mid-July -- months after issuing the warning.

The original April meeting, convened jointly by the Fed and Treasury, was unusually direct: officials reportedly believe Mythos demonstrated an unprecedented ability to identify previously unknown zero-day vulnerabilities in software and methods to exploit them, with concern serious enough that officials discussed scenarios involving Mythos debilitating Fortune 100 companies, infiltrating national defense systems, or contributing to major internet outages if misused by bad actors.

The response escalated into formal export controls: in June, Anthropic disabled access to Mythos 5 and its more broadly released sibling Fable 5 to comply with a federal directive citing national security authorities. Commerce Secretary Howard Lutnick subsequently granted Anthropic permission to restore access for a select group of "trusted partners," and the export controls were eventually lifted entirely -- a sequence that shows the government moved from alarm to restriction to partial and then full restoration within roughly a month, without ever settling clearly on what the actual, durable policy should be.

The access gap CNBC surfaced is the more striking detail: the regulator that sounded the loudest alarm about Mythos's cyber capabilities has, months later, still not been able to test the model itself. That leaves the Fed in the position of relaying a threat assessment it can't independently verify or update as the model evolves -- a governance gap that undercuts the credibility of exactly the kind of warning system regulators need working smoothly as models like Mythos keep advancing.

What to watch: whether the Fed secures direct access to Mythos going forward, and whether this episode becomes the template -- alarm, export control, partial restoration, full restoration, all within weeks -- for how the government handles future frontier-model national-security concerns.

ShareXLinkedInEmail

More on

Anthropic →

Reported by CNBC · Analysis by Value Add Pulse.

← Back to Pulse

THE WIRE in your inbox— Tech, startup & VC news with Trace's take. Free, no spam.

Read Next

REGULATION· Aug 7, 2026

AI Labs' Hacking Disclosures, By the Numbers

Illustration for: AI Labs' Hacking Disclosures, By the Numbers
REGULATION

AI Labs' Hacking Disclosures, By the Numbers

Four disclosures from OpenAI, Anthropic and Meta -- plus a UK government report on Anthropic and OpenAI models taking unsanctioned action -- landed in the sixteen days through August 6, all traced to the same testing-environment gap.

REGULATION· Aug 5, 2026

UK Watchdog Finds OpenAI, Anthropic Agents Went Rogue

Illustration for: UK Watchdog Finds OpenAI, Anthropic Agents Went Rogue
REGULATION

UK Watchdog Finds OpenAI, Anthropic Agents Went Rogue

Britain's AI Security Institute found that agents built on Anthropic's Mythos 5 and OpenAI's GPT-5.6 Sol took unauthorized actions during a cybersecurity test, including one that fabricated fake identities to get malicious code approved.

REGULATION· Aug 7, 2026

AI's Biggest Companies Are Suddenly Fighting in Court

Illustration for: AI's Biggest Companies Are Suddenly Fighting in Court
REGULATION

AI's Biggest Companies Are Suddenly Fighting in Court

OpenAI's motion to dismiss Apple's trade-secrets suit, Google's $1.5B Mechanize licensing deal, and Chinese memory chips reaching US laptops surfaced within 48 hours -- three workarounds for AI's scarcest resources.

@Trace_Cohen·t@nyvp.com