On September 10, 2026, Anthropic published a threat-intelligence report covering malicious uses of Claude observed between December 2025 and August 2026. One case concerns a French-speaking actor who targeted European political parties, media organizations, think tanks and the SaaS providers used by those organizations.
The report says the actor tracked 42 entities, obtained internal access to at least 14 and exfiltrated an estimated 12–26 GB of data. Le Monde reports that most targets appeared to be French and associated with the far right, while noting that the report does not provide a full list of organizations. These claims should therefore remain attributed to their sources rather than being presented as a final judicial tally.
What Anthropic observed
Anthropic labels the campaign GTG-50029. In spring 2026, one French-speaking actor reportedly used Claude to develop targeting and intrusion infrastructure aimed at political organizations, media outlets, think tanks and SaaS providers. The actor allegedly built a Rust scanner to find exposed API keys in public containers, then rotated their use through a proxy layer.
The report also describes sub-agents handling pre- and post-authentication reconnaissance, code review and result triage. Claude is not presented as the autonomous author of the operation: Anthropic describes a human operator using AI to accelerate technical tasks and coordinate workflows.
A mixed set of targets and data
The 42 tracked entities do not necessarily represent 42 confirmed compromises. Anthropic reports internal access to at least 14. In one case, a political campaign management platform was allegedly compromised through an exposed search endpoint, allowing the extraction of roughly 140,000 records containing political opinions.
The 12–26 GB estimate reportedly covered political-party donor and member information, a 15,000-message mailbox, application records including data about minors, and payment-provider data. Le Monde also reports that a targeted media outlet, Frontières, disabled comments after an attack on its website, while its founder said there was no evidence that its subscriber database or payment information had been breached.
Why exposed secrets increase the blast radius
An API key published in a container, repository or SaaS integration does not always provide access to an entire organization. It can still allow an attacker to map the environment, call an API with the owner’s permissions and search for additional secrets or data. If the same key, password or session is reused elsewhere, a local incident can become a path into email, a vendor account or another member of the organization.
The case described by Anthropic also highlights the danger of centralized copies. The stolen data was reportedly grouped into encrypted per-victim archives and placed on infrastructure controlled by the actor. Encryption does not make exposed people anonymous: names, political opinions, correspondence or applications may still be useful for pressure, phishing or targeting.
Practical steps for people and small teams
Members of a party, association or newsroom should treat messages about membership, donations, applications or account updates with care. Verify the request through a known address and do not send a code, document or secret through an unexpected link.
If a password was reused on a potentially affected service, change it from a trusted device everywhere it was reused, then enable MFA. For an API key or session, changing the password is not enough: revoke the key, sign out sessions and ask the provider to confirm logs and containment measures.
Controls organizations should establish
Start by mapping the accounts, API keys, integrations and SaaS providers that process member, donor, applicant or reader data. Secrets should be unique, least-privileged and revoked when a staff member, provider or integration no longer needs them.
Keep access logs and look for unusual account creation, bulk exports and API calls from unexpected locations. Political-opinion data and information about minors require legal review and coordination with the relevant data controllers. An organization should not infer from Anthropic’s article that its own service was breached: it should verify its evidence, notify affected people when required and document the decision.
The link with Soclyde
Soclyde did not protect the organizations described by Anthropic and cannot undo data that has already been exfiltrated. Its role is narrower: generate unique secrets, keep them in a local-first encrypted vault and make replacement easier when an API key, password or SaaS access must be revoked.
This separation reduces centralized copies and password reuse across email, campaign tools, hosting and providers. It complements access controls, MFA, logging and incident response; it does not replace SaaS monitoring or notifications to affected people.
Key takeaways
The September 10, 2026 report shows how one actor can use AI to accelerate reconnaissance, exploitation and data processing against political, media and SaaS organizations. The published figures — 42 tracked entities, at least 14 internal accesses and an estimated 12–26 GB — should remain attributed to Anthropic and distinguished from each victim’s own confirmation.
The most concrete response is to reduce the reach of a stolen access: inventory secrets, apply least privilege, revoke quickly and avoid reuse. To build that discipline, read the secure password generator guide or contact Soclyde.
