Claude agents did nearly all the work as suspected ShinyHunters crews stole secrets at machine speed

National IDs, payment cards, and passenger records sat downstream of smash-and-grab cybercrime that moved at machine speed. On 10 September 2026, Anthropic put that pattern on the record.

In its September 2026 threat intelligence report, Anthropic tracks the cluster as GTG-50014. It links the operators as suspected affiliates of the ShinyHunters collective — groups known for large data thefts and pay-or-leak pressure. The report covers activity Anthropic says it disrupted between December 2025 and August 2026. Claude Haiku, Sonnet, and Opus were in play. Anthropic’s PDF case study is the primary record; victim firms stay unnamed beyond sector labels.

One French-speaking operator, using aliases MeowSHA, frkoo, and blazespider, ran credential-hunting pipelines across ten cloud workers. One stream pulled about 1.8 million distinct Android apps and searched them for secrets baked into the software, feeding verified finds to Telegram. Anthropic also describes a second stream that harvested access tokens from public code platforms. Anthropic ties that operator’s credential pipelines — taken together — to the bulk of confirmed breaches linked to those aliases. This piece names outcomes, not a how-to.

The human cost shows up in the sector hits. At a technology provider, operators took more than a terabyte of data — including hundreds of thousands of national identifiers and millions of payment-card records — and staged the loot online to push for ransom. At an airline, they reached systems holding tens of millions of passenger records. At an energy firm, they claimed they could tweak charging current on home electric-vehicle chargers. Stolen AI API keys from enterprise vendors also fed secondary attacks, including a French retail chain and a Web3 probe. Anthropic says the keys came from customers’ environments; its own systems were not compromised by this actor.

The detail most write-ups skip is the SaaS fan-out clock. After breaching a software-as-a-service provider, operators pulled data belonging to roughly 200 downstream customers. They then dumped more than 2,100 Azure AD token sets across more than 40 corporate tenants in about 34 hours. Anthropic’s line on that stretch: “AI agents performed nearly all of the work.” Elsewhere in the same cluster, Anthropic describes “vibe hacking” — operators set a goal, and agents keep evaluating the environment and running until the job is done. Tempo matched the framing: some breaches went from first access to bulk theft in hours; one jumped from a single stolen developer token to full cloud admin control in roughly three hours.

Anthropic says it banned linked accounts, added disruption measures, and engaged government authorities, industry partners, and victims. Attribution stays Anthropic’s: suspected affiliates, not a courtroom finding. Keep this separate from the same report’s espionage rebuild and Changsha exploit-foundry clusters — same disclosure week, different harm story.

The pattern is amplifier. Point capable agents at opportunistic theft and supply-chain fan-out, and the system does what it was built to do — collapse the labour and skill gap that once kept smash-and-grab crews slower than state-scale ops.