China-linked spies impersonate Anthropic to spy on the U.S. debate over regulating AI

By Zendoric · October 5, 2026.

A hacking group that cybersecurity firm Proofpoint links to the Chinese government, dubbed "TA419," has impersonated Anthropic researchers and a former senior White House official to deceive artificial intelligence policy specialists at U.S. think tanks, universities and law firms. The target, according to Proofpoint's report, was the cloud accounts of the people writing the rules for American AI.

The modus operandi was patient. The attackers opened contact with innocuous-looking emails: an invitation to join a nonexistent "AI Policy Advisory Committee" or to contribute to a supposed report for the Senate Foreign Relations Committee. Only after several exchanges that built trust did they send malware-laden documents designed to gain access to their victims' accounts. In one documented case, someone posing as a senior Anthropic executive used the subject line "Request for Feedback on Military Integration of Claude" to approach a person working at a think tank.

Among those impersonated is Lynne Parker, who served as principal deputy director of the White House Office of Science and Technology Policy during the Biden administration. Parker told CNN that, beyond her concern for her colleagues, she felt "sadness" at seeing relationships built over the course of her career exploited by malicious actors to deceive others. Mark Kelly, a Proofpoint researcher, told the network that the firm has "confidence" that this is an actor aligned with the Chinese government, based on the pattern of targets, the infrastructure observed and corroboration from other companies in the sector; it is worth stressing that this is a technical attribution by a private firm, not a formal accusation by any government.

The episode comes at a particularly charged moment in the U.S. regulatory debate on AI. Dario Amodei, Anthropic's chief executive, is one of the executives calling for rules for an industry that, they warn, could produce anything from security crises to catastrophic scenarios if it develops without oversight. At the other end, the White House and figures such as Mark Zuckerberg (Meta) and Jensen Huang (Nvidia) oppose such regulation. This very week, the Trump administration hosted a summit at which several companies signed a voluntary "self-regulation" agreement, which the president himself called "morally binding." One of the core arguments of those who reject tougher rules is that the U.S. cannot afford to cede ground to China in the AI race.

And there, in our view, lies the central irony of this case: the very arms-race logic invoked to hold off regulation is precisely what turns the regulatory debate into an espionage target. If the big geopolitical question is who sets the rules for AI, then knowing in advance which way Washington is leaning, what arguments are circulating among Senate advisers or what the sector's major companies think internally is as valuable as a stolen model. The target was not a lab; it was the fabric of trust among the experts who advise lawmakers.

This fits a reading we have been advancing at Zendoric on AI safety: the most relevant risk indicators are not saturated offensive-capability benchmarks, but social engineering aimed at specific institutions. No sophisticated automated cyberattack model was needed here: it was enough to impersonate a trusted company and have the patience to cultivate a relationship before launching the malware. The risk lies not only in what AI can do, but in how the prestige and brands of the companies building it become a weapon for anyone who wants to spy on the process that decides how that same technology is governed.

In the short term, the most likely effect is a chilling of open collaboration among experts, think tanks and AI labs: every email from a supposed Anthropic, OpenAI or Google researcher will now have to pass through an extra layer of verification, which slows precisely the kind of exchange that produces good public policy. In the long term, episodes like this are a reminder that no society reaches the abundance AI promises without first solving the most basic thing: knowing with certainty who is on the other end of an email. Digital trust infrastructure — robust authentication, identity verification, institutional security hygiene — is not an obstacle to AI progress; it is its foundation.