Anthropic Opens Mythos 5 to EU Regulators, Lands Andrew Tulloch

Three developments in the past 48 hours show Anthropic simultaneously courting regulators, absorbing top talent, and managing a safety narrative that keeps growing more complicated ahead of its anticipated IPO.

Key points

• The EU’s cybersecurity agency has been granted direct access to Mythos 5, Anthropic’s model capable of detecting software vulnerabilities, to assess risks and safeguards across member states.
• The regulatory access follows national-security concerns that Mythos 5 has already raised, making the EU’s evaluation politically significant.
• Star AI researcher Andrew Tulloch has joined Anthropic from Meta, where he had received a reported billion-dollar compensation offer less than a year ago.
• Separately, an internal alignment researcher publicly estimated a greater-than-10% extinction risk from Anthropic’s technology, and a colleague resigned, compounding investor-relations challenges for the upcoming IPO road show.

Why handing Mythos 5 to EU regulators matters

Anthropic’s decision to give the European Union Agency for Cybersecurity access to Mythos 5 is notable for what the model actually does. According to the Wall Street Journal, it can detect software vulnerabilities, a capability that has already raised national-security flags. Voluntarily opening that system to regulatory testing is a calculated move: it positions Anthropic as cooperative rather than evasive at a moment when European AI oversight is tightening.

For investors watching the IPO timeline, the EU access agreement cuts two ways. It could accelerate regulatory clearance and reduce the risk of a sudden enforcement action in a major market. But it also formally confirms that Mythos 5 sits in a sensitive enough category to warrant government-level scrutiny. Any findings the EU agency publishes, or leaks, could affect how institutional investors price regulatory risk into the offering.

Can Andrew Tulloch change Anthropic’s research trajectory?

The Tulloch hire, reported by the Wall Street Journal, is a striking talent coup. He departed Meta despite a reported billion-dollar compensation package received less than a year ago. The details of what Anthropic offered are not public, but the willingness to leave that arrangement signals either an unusually competitive counter-offer, a strong preference for Anthropic’s research environment, or both.

The broader context matters here. Anthropic has consistently argued that the best way to make powerful AI safe is to have the best researchers building it. Tulloch’s arrival reinforces that pitch. It also puts pressure on Meta and other labs that assumed headline compensation numbers would be sufficient retention tools. For IPO purposes, landing a researcher of this profile supports the narrative that Anthropic can attract elite talent even as it competes against better-capitalized rivals.

How bad is the internal safety dissent problem?

This is the harder question. The Information reported that an Anthropic alignment researcher publicized a greater-than-10% extinction risk estimate tied to the company’s own technology, and that a colleague subsequently resigned. The timing, overlapping with IPO road show preparations and political outreach, is as damaging as the content.

The specific concern for prospective investors is not abstract. Public statements about extinction-level risk from a sitting Anthropic researcher carry a different weight than third-party criticism. They become potential disclosure material, fuel for congressional scrutiny, and fodder for plaintiffs’ attorneys if the IPO proceeds and shares later decline. The resignation amplifies the signal that the dissent is not a single outlier view.

Anthropic’s challenge is that its founding story is built on safety-first credibility. That same credibility means internal safety researchers speaking out cannot simply be dismissed as uninformed critics. The company will need to address these concerns substantively during the road show rather than deflect them, which is a harder and slower task than managing external commentary.

Taken together, the three stories from this 48-hour window trace a company that is simultaneously making the right regulatory gestures, winning on talent, and facing a safety-narrative problem with no quick fix. All three threads will be live issues when the IPO prospectus lands.