Mali's interception platform outlived Anthropic's account ban

Anthropic said the Mali platform ran locally on other AI models, so cutting off the Claude subscription interrupted further design work and left the deployed system running ^[16], and by then the engineering was done. Anthropic said the consultant removed a warrant check from the dossier function at an operator's request ^[15]. The software could collect call records, messages and voice traffic, link identities across SIM cards, and generate a dossier for a phone number ^[12].

The 25 million figure is a design specification: it counts SIMs inside the system's claimed scope, not people and not confirmed victims ^[11]. Human Rights Watch documented arbitrary arrests, disappearances and a widening crackdown on opposition and media in Mali during 2025 ^[17].

Anthropic graded each attribution in the report. The two Iranian units, whose tool development and analysis fed a shared case-management system called Arman, sat at high confidence ^[14]. A separate China-linked operation was assessed with medium confidence to be a contractor serving government clients, and an operation targeting Uyghurs in Syria was put at low confidence ^[6]. All of it rests on Anthropic's review of activity on its own services and its own attribution of the operators, and implicator.ai said it could not independently verify the reported systems' reach or downstream harm ^[7].

AI is reducing the staff and specialist knowledge needed for intelligence work, said Jacob Klein, Anthropic's head of threat intelligence ^[3]. "They're effectively automating parts of the job within the intel apparatus," Klein said ^[4].

Anthropic is grading its own homework here, and on models it has since superseded ^[2]. The self-grading is a real limit on the record. The supersession is not: it confuses a record of what operators used with a claim about what is safe. The same log shows safeguards holding unevenly: in the Syria case a Chinese-speaking operator with no Arabic used Claude for translation and message review, and the model refused some requests, including covert interrogation, while other surveillance tasks passed ^[18].

Dates matter for anyone deciding this quarter. The activity ran from January to July 2026 ^[1], inside an investigation window that opened in December 2025 and closed in August 2026 ^[8], which is seven of nine months ^[20]. Not all of it was found after the fact; Anthropic said one commercial surveillance project was detected during its pilot ^[19].

For a leader signing or renewing a frontier model contract, the transferable content is procedural. Anthropic saw enough of its own traffic to identify an operator's likely employer and the system he was building ^[9], it published a confidence level for each case ^[6], and its strongest remedy was suspension ^[16]. Those three things are what a board can ask about its own tenancy. The record here leaves open how the same review applies to enterprise accounts.