Skip to content

Google GTG

Summary

At LABScon on September 18, 2026, Google Threat Intelligence Group (GTG) researcher Austin Larsen publicly detailed, for the first time, that a Mandiant undercover analyst had been inside TeamPCP's inner circle "from almost day one" of the group's 2026 supply-chain rampage. A Google persona that had spent months building trust with an actor later invited into TeamPCP was added along with him to the group's core chat, one of about 12 members of the channel the group called CanisterWorm, giving Google a live view of the credential vault, the extortion planning, and the infighting while the Trivy → LiteLLM → TanStack/Mistral-era compromises were still unfolding. The talk (reported by WIRED) supplies the first end-to-end public narrative connecting the August 27 AFP/WAPF/FBI charges to Google's investigative work: leaked BreachForums data tied one of the chat's most-active handles to sheepstealing@gmail.com, a 2019 pirated-Office-key refund dispute tied that handle to ruben@thomsonfamily.net.au, i.e., accused principal Ruben Ian Thomson, and the clincher was TeamPCP backing its relocated stolen-credential server up to a Google Drive owned by that same Gmail account, which Larsen tipped to the FBI ("an agent responded in a matter of minutes"). The talk also produces two new durable intelligence items: (1) ShinyHunters partnered with TeamPCP around April 2026, then went rogue, extorting victims with TeamPCP's own stolen credentials while withholding TeamPCP's cut, even volunteering its chat log from TeamPCP's server to Larsen, prompting TeamPCP to exile ShinyHunters (and Google's mole) and move its data; and (2) someone inside the core circle was using an AI tool to develop a zero-day exploit against a widely used login product to bypass its two-factor authentication, Google obtained the exploit code, confirmed it worked after a few tweaks, and warned the vendor, which patched (this is the incident Google described in an anonymous May 2026 case study, now, for the first time, placed inside TeamPCP). Monetization math is also new: despite the AFP's 500,000+ stolen credentials, Larsen says TeamPCP was extracting only tens of thousands of dollars in extortion, not millions, which is why it brought revenue-share partners in, the decision that produced the betrayal. Google's disruption model is the closing thesis: rather than notify hundreds of victim companies one by one, GTG went to credential-issuing platforms first (AWS, Microsoft) to have stolen credentials revoked before they could be used, sending hundreds of provider notification emails, "writing reports can only be so useful," under the newly formed Google Cyber Disruption Unit.

Tags

What was disclosed (LABScon, Austin Larsen, Sep 18, 2026)

The mole

  • A Google/Mandiant undercover persona had spent "many months building trust with one of the actors that was invited to join TeamPCP, and so was added to the group." Effectively "almost day one, Mandiant was watching everything behind the scenes." The analyst's name was not revealed.
  • The inner circle was ~12 members with access to a core chat the group named CanisterWorm (note: on this wiki the name also tracks the npm worm of the same name, the chat and the worm shared the group's branding).
  • Access included the server where TeamPCP stored its trove of stolen usernames, passwords, and access tokens, the material it planned to extort with (AFP figure: more than half a million users' credentials).
  • Outside validation of the mole's existence: Telstra threat researcher and former AFP analyst Michael Fletcher says he approached Larsen around that time about monitoring TeamPCP and was told to approach the hackers carefully because one of them was "friendly."
  • Guardrails: Larsen says the analyst never engaged in illegal hacking or encouraged breaches, "a fly on the wall, only saying enough to not be suspicious."

The money problem and the ShinyHunters betrayal

  • Despite the credential haul, Larsen estimates TeamPCP pulled in only tens of thousands of dollars in extortion payments, versus the millions similar groups amass. Struggling to monetize, TeamPCP invited multiple cybercriminal groups in as revenue-share partners, granting them access to the stolen credentials for a percentage of what they extracted.
  • ShinyHunters joined around April 2026. "A few weeks" later it went rogue: running its own extortions with TeamPCP's credentials and withholding TeamPCP's cut, and unsolicited, shared with Larsen a full log of its own chat on TeamPCP's server, not knowing Google already had access via the mole. ShinyHunters also taunted TeamPCP on X.
  • TeamPCP's response: narrowed the inner circle, moved its data to a server at a different hosting provider, and exiled ShinyHunters, and several other members, including Google's undercover analyst, from CanisterWorm ("Just delete that and stop sharing shit with shinyhunters," one leader wrote).
  • Convergence: Google's account independently corroborates the Anthropic September 2026 report's GTG-50014 cluster ("suspected ShinyHunters affiliates") running AI-assisted extortion at industrial scale in the same months, and extends the ShinyHunters pattern seen in Unit 42 token-jacking and the Salesforce/OAuth abuse reporting.

Disruption before notification

  • With vault visibility, Larsen's team chose disruption over sequential victim notification: first contacted platforms where the stolen credentials worked, AWS and Microsoft, to get credentials revoked en masse, then sent hundreds of notification emails to providers and victims. "Let's go mess up what they're doing. That was my goal."
  • The in-the-wild AI-built zero-day: via the inner-chat visibility, Google learned someone in the core circle (separate from the supply-chain ops) was using an AI tool to develop a zero-day against a widely used login product to bypass 2FA. Google got a copy of the exploit code, tested it, found it worked with a few tweaks, and warned the vendor, who patched. Google had described this in a May 2026 case study without naming TeamPCP, the LABScon talk is the first public placement of it inside the group. One of the rare publicly confirmed instances of an AI-developed, previously unknown vulnerability exploited in the wild.

The Opsec trail to the arrests

  • BreachForums leak: one of the most active handles in CanisterWorm had registered on the forum with sheepstealing@gmail.com.
  • 2019 forum archive: a refund dispute between user "sheepstealing" and a pirated-Microsoft-Office-key seller demanded the refund at a PayPal account tied to ruben@thomsonfamily.net.au → identity pivot to Ruben Ian Thomson (charged Aug 27, 2026, Perth Magistrates Court, with Louis Michael Gaebler).
  • The clincher: after the post-betrayal server move, Google learned (via a "trusted partner") some contents of the new server and that it was being backed up to a Google Drive on that same sheepstealing@gmail.com account, stolen material synced to an account tied to the operator himself. Larsen tipped the FBI; an agent responded within minutes; ~a month later law enforcement completed warrant requests to Google for Thomson's data; the arrests followed in late August.
  • Brian Krebs independently published his own clue trail to Thomson's identity the prior month; Google's tip was not necessarily the only path.

Attribution discipline and framing

  • Leaked-chat boast: "You guys should understand that we pulled off the biggest supplychain [sic] maybe ever recorded in modern history."
  • WIRED's recap of the campaign list matches this wiki's record: Trivy, LiteLLM, Checkmarx infrastructure, TanStack, Mistral AI, then breaches of GitHub, Mercor, and employee devices at OpenAI, the European Commission, and others unnamed; Mini Shai-Hulud as the automation worm, named after the Sep 2025 original whose lineage to TeamPCP remains unestablished.
  • Larsen frames the operation under Google's new Cyber Disruption Unit: "Writing reports can only be so useful. Taking action to protect users and customers, that is the next step."

Durable defender reads

  • Revocation-beats-notification ordering is a reusable IR playbook shape for mass credential theft: at hundreds-of-victims scale, engaging credential issuers (cloud providers, IdPs) to revoke class-by-class precedes per-victim outreach and cuts the attacker's use-window.
  • Criminal-partnership instability is an intelligence source and an operational risk: revenue-share with prolific groups (ShinyHunters) traded monetization reach for betrayal risk; expect stolen-credential ecosystems to be contested, credentials you rotate may already be in the hands of the thief's ex-partners.
  • The AI-developed 2FA-bypass zero-day confirms agents-in-the-loop exploit development is not hypothetical: patch urgency for authentication software should assume adversary AI tooling is fuzzing/repair-looping it.
  • Person-level Opsec, not infra, ended the operation (forum handle → email → PayPal → family domain → personal Google Drive). For defenders the mirror-image lesson: attacker-attribution trails come from identity plumbing, and law-enforcement outcomes do not close credential exposure, rotation across the March 2026 exposure window still stands.

Confidence and caveats

  • Single primary source (Larsen's LABScon talk as reported by WIRED); Google has not published a written report of the details at disclosure time. The talk text/deck may add or correct detail.
  • "About 12 members," monetization estimates ("tens of thousands"), and the timing of the ShinyHunters partnership come from Google's account of its own investigation, unfalsifiable from public data; treat scale and interior-chat details as one-party claims.
  • Charged individuals are alleged, presumed innocent; the underlying charges are covered on the AFP/WAPF/FBI charging page.
  • Google/Mandiant has its own institutional interest in the disruption narrative; the "fly on the wall" characterization is self-reported.
  • WIRED names the core chat "CanisterWorm" as TeamPCP's own label, consistent with, but not identical to, the wiki's CanisterWorm worm naming; do not conflate the chat with the npm worm artifact.

Sources