An unverified account on X claims that Anthropic’s Claude Code helped identify hidden spyware on a woman’s phone, finding an app allegedly transmitting her location every two minutes for 11 weeks. The report has not been independently confirmed, but it illustrates how AI agents could make technical security investigations more accessible to non-specialists.
According to the account, the suspected app was called “System Service,” had no visible icon and was not installed through an official app store. Claude Code reportedly catalogued installed software, permissions and network connections, then helped preserve evidence before the app was removed.
- Hidden apps or software installed from unknown sources
- MDM profiles and administrator privileges
- Persistent access to location, microphone or camera
- Processes configured to launch automatically
- Connections to unfamiliar servers or IP addresses
The key change is the interface: rather than requiring someone to interpret a conventional antivirus report, an AI agent can be asked to investigate a symptom and guide the examination. The same X account also cited a separate instance in which Claude Code reportedly found a cryptocurrency miner on a Mac after an unknown process was consuming nearly 12 CPU cores.
Security vendors are already adding conversational and automated layers to their products. Bitdefender’s Scamio analyzes messages, links and QR codes in a chat-style interface, while SentinelOne’s Purple AI is designed to investigate alerts and initiate automated responses.
That shift does not make established detection engines, threat telemetry or antivirus databases obsolete. It could, however, move the user relationship toward general-purpose agents that call security tools in the background rather than standalone scan applications.
The capability also has an offensive side. Anthropic has documented misuse of Claude for data theft, extortion and cyberespionage, including an operation targeting at least 17 organizations. AI agents may lower the technical barrier both for people investigating suspicious activity and for attackers seeking to automate it; the unresolved question is how reliably those agents can distinguish a real compromise from a false alarm before taking action.
