// 1 ZERO-DAY · 2 CVE · 2 EXPLOIT IN THE LAST 24H→
The China-aligned espionage group TA419, tracked by Proofpoint, has conducted targeted phishing campaigns against AI policy experts in the United States and Japan. Active since at least April 2025, with spikes in February and July 2026, the operation marks a shift toward the AI "policy layer"—the people writing the rules—using a refined adversary-in-the-middle attack built on a modified version of the open-source Frameless BitB tool to bypass multi-factor authentication and capture live session tokens.

The China-aligned cyber espionage group TA419, tracked by Proofpoint, has conducted targeted phishing campaigns against artificial intelligence policy experts in the United States and Japan. The activity is documented from at least April 2025, with spikes in February and July 2026, and marks a shift in targeting toward the so-called "policy layer" of AI: not the models, not the training data, but the people who write the rules. The technical mechanism is refined: an adversary-in-the-middle attack based on a modified version of Frameless BitB, an open-source tool, that bypasses two-factor authentication by capturing credentials and active sessions.

Key Takeaways
  • TA419 impersonated real institutional figures: Lynne Parker, former Principal Deputy Director of the White House OSTP; economist Heidi Crebo-Rediker; and an unnamed senior Anthropic employee.
  • The attack uses an AitM reverse proxy with modified Frameless BitB: a fake browser window that relays the real Microsoft authentication, capturing credentials and post-MFA session tokens.
  • Proofpoint estimates fewer than 10 individuals across a handful of organizations were targeted in the July 2026 campaign; it is not confirmed that any accounts were actually compromised.
  • Proofpoint researcher Mark Kelly bases the Chinese alignment on "targeting consistently in line with Chinese government interests, observed infrastructure and technical artifacts, and corroboration from industry partners."

The Stolen Identities: From Lynne Parker to the Senior Anthropic Employee

In July 2026, TA419 sent emails posing as Lynne Parker, former Principal Deputy Director of the White House Office of Science and Technology Policy. The emails, sent starting July 8 according to Proofpoint, invited recipients to a fictitious "AI Policy Advisory Committee." The group subsequently impersonated Heidi Crebo-Rediker, an economist and foreign policy expert. Earlier, in February 2026, TA419 had used the identity of a senior Anthropic employee to contact an AI policy analyst at a U.S. think tank, with the subject line "Request for Feedback on Military Integration of Claude."

Parker confirmed that two people received emails in her name. Parker, quoted by the Times of India, stated: "My first concern was for the colleagues who might receive the impersonated email. I wanted to warn them but realized that's very difficult to do when I don't know who is being targeted." This detail reveals a structural characteristic of the attack: the impossibility for the impersonated figure to alert potential targets, whom they do not know.

How the AitM Mechanism Works with Frameless BitB

After the target's initial reply, TA419 sent a shortened link with multi-stage redirects to a fake OneDrive page. The technical core is an adversary-in-the-middle reverse proxy based on a modified version of Frameless BitB, an open-source tool available on GitHub. Frameless BitB creates a fake browser window without iframes, faithfully mimicking the Microsoft login.

The difference from traditional phishing lies in the relay: the proxy intercepts the interaction with the real Microsoft authentication, capturing both cleartext credentials and the session token generated after MFA completion. Proofpoint documents that TA419 added a custom module for three specific functions: monitoring login progress, automatically checking the "Keep me signed in" option, and injecting OTP codes upon acceptance. The result is an automatic bypass of SMS- or app-based two-factor authentication, with persistence of the captured session.

The tool does not exploit a Microsoft vulnerability: it is social engineering channeled through a proxy that exploits the user's trust in the authentic interface. The target sees the real login, completed successfully, without realizing the entire interaction has been mediated.

The Policy Layer as the Objective: Not Stealing Models, Anticipating Rules

Proofpoint assesses that the campaigns feed Chinese intelligence gathering on the development of AI policy and regulation in the United States. The targeting is narrow and vertical: think tanks, universities, law firms, with specific pretexts such as a Senate Foreign Relations Committee report on export controls and the AI supply chain. This pattern distinguishes TA419 from actors targeting technical IP exfiltration or industrial data.

The objective is to learn regulatory directions, export restrictions, and national security criteria applied to AI models in advance. The technical distance is high: an AitM with a custom module and automatic MFA bypass. The human proximity is equally sophisticated: impersonation of known figures, plausible pretexts in the policy-specific context, timing coherent with the public debate. According to Reuters, citing Proofpoint, fewer than 10 people in a handful of organizations were targeted. The source does not specify whether lookalike domains resembling the Heritage Foundation, World Economic Forum, or Japan-Taiwan Exchange Association were used in this specific campaign or in previous activity.

"Individual targets in scope of TA419 activity should treat unsolicited subject-matter outreach as a plausible pretext stage, and seek to verify the legitimacy of such unexpected communications via another independent medium." — Proofpoint

What to Do Now

  • Verify via an independent channel any unexpected outreach on policy topics, even if it appears to come from known figures: the Proofpoint brief explicitly identifies this as the first control.
  • Recognize that traditional SMS and OTP apps are not sufficient against AitM of this caliber: the authentication succeeds on the real server, rendering the proxy invisible.
  • Treat shortened links with multi-stage redirects as high-risk indicators, especially if they follow an email conversation initiated on an institutional pretext.
  • Consider that the AI "policy layer"—those writing rules, export controls, and national security criteria—is now a documented target of Chinese espionage, not just an academic or institutional domain.

The Limit of the Case: Between Attempt and Compromise

The dossier has a hard boundary. Proofpoint documents the attempts, the infrastructure, the tools, and the impersonated identities. It does not confirm that accounts were actually compromised or that emails were exfiltrated. This limit does not diminish the relevance of the attack: the technical structure demonstrates advanced operational capability, and the vertical targeting of AI policymakers indicates strategic intelligence gathering, not opportunism. It does not document specific remedial measures beyond cross-channel verification.

The emerging picture is of an espionage ecosystem that has internalized the structure of the American AI debate: who matters, what they discuss, how they contact each other. There is no need to breach Anthropic or the White House if you can induce an analyst to believe they are speaking with someone who worked there.

FAQ

What is Frameless BitB and how does it differ from traditional phishing?

It is an open-source tool that creates a fake browser window without iframes, mimicking the login of real services. Unlike classic phishing, which presents a fake page and captures only credentials, the AitM version used by TA419 relays the interaction with the real authentication, capturing the post-MFA session token as well.

Why didn't two-factor authentication stop the attack?

SMS- or app-based MFA is completed successfully on the authentic server: TA419's proxy forwards the code in real time. The session token generated after valid authentication is intercepted, allowing the attacker to impersonate the user without needing to re-authenticate.

Were Lynne Parker or Anthropic breached?

No. Parker and the Anthropic employee were impersonated, not compromised. Their identities were used as a pretext to contact third parties. Anthropic as a company was not breached; the attack exploited the identity of a person who works there.



Sources

Sources and references
  1. cyberscoop.com
  2. infosecurity-magazine.com
  3. lineaedp.it
  4. findarticles.com
  5. timesofindia.indiatimes.com
  6. fedscoop.com
  7. github.com