Anthropic Says State-Linked and Criminal Actors Tried to Weaponize Claude for Android App Secret Extraction

According to a report from BleepingComputer, Anthropic disclosed that multiple threat groups attempted to abuse its Claude AI model for malicious purposes, including financially motivated criminal actors and state-sponsored espionage groups reportedly linked to Russia and China. Per the report, the abuse involved attempts to use Claude to extract secrets from a large corpus of roughly 1.8 million Android applications.

Anthropic is the source for these claims, and the reporting does not provide granular technical detail on the exact extraction methodology, what categories of secrets were targeted (e.g., API keys, hardcoded credentials, embedded tokens), or which specific actors were confirmed responsible for which activity. Treat the attribution to Russia- and China-linked groups as Anthropic's characterization pending further corroboration.

Why It Matters for Defenders

This report is another data point in a broader trend: large language models are increasingly being used to accelerate reconnaissance and secret-hunting at scale against mobile application ecosystems. If threat actors can use AI assistance to rapidly triage millions of APKs for embedded credentials, cloud keys, or hardcoded tokens, the effective cost of large-scale secret-scraping campaigns drops significantly. Organizations that ship Android apps — and any organization whose secrets could end up embedded in third-party or first-party mobile binaries — are potentially affected, regardless of whether they were a specific named target.

What Defenders Should Watch For

• Audit your own Android app releases (and any SDKs/libraries you ship) for hardcoded secrets, API keys, or tokens using standard secret-scanning tooling as part of your build/release pipeline.
• Treat any credential or key ever embedded in a shipped mobile binary as potentially compromised, since decompilation and secret extraction from APKs is not new — AI assistance mainly changes the speed and scale of this activity.
• Monitor for anomalous use of exposed keys/tokens (unexpected API call volume, logins from new regions or ASN ranges, unusual service usage patterns) as a downstream indicator that embedded secrets have been harvested and are being used.
• Review vendor and AI-provider abuse disclosures periodically — providers like Anthropic publishing threat intelligence on abuse of their own models is a useful, relatively new source of early warning for this class of activity.

This is developing, third-party-sourced intelligence with limited technical detail published so far, and df00tech has not independently verified the underlying claims. For the original reporting, see BleepingComputer's coverage.