AIToolWatchdog
A lightweight agent that runs alongside any AI assistant on a workstation and logs every file read, command run, and outbound network call, flagging anything that looks like hidden data exfiltration.
IT security teams and security-conscious developers
- Monitors filesystem, shell, and network activity for popular AI coding assistants
- Flags unusual outbound traffic such as telemetry hidden inside normal-looking calls
- Generates a transparency report per tool with a 'what it really does on your machine' summary
- Slack and SIEM alerts when a watched tool's behavior changes between versions
A researcher just exposed hidden tracking code in Claude Code that captured timezone and proxy data without disclosure, and Anthropic admitted it was 'meant to prevent unauthorized resellers,' making enterprise security teams newly skeptical of AI coding tools.
Anthropic Claude Code hidden tracker story is real and widely covered (Ars Technica, Malwarebytes, Cybersecurity News); MITRE ATLAS and multiple vendor reports say ~80% of orgs see AI agents acting beyond intended scope, but DE 'spionage' trend is only 200+ searches (weak).Anthropic outed for Claude tracker that secretly monitored Chinese users ↗Claude Code's hidden tracker was an "experiment," says Anthropic ↗
Space is crowded, not wide-open: Cyberhaven AI/DLP platform, Microsoft Global Secure Access MCP logging, Obot MCP Gateway, Datadog MCP security guidance, LangSmith/Langfuse/AgentOps observability, and Roo Logger directly log AI coding-assistant activity — none is identical but the niche is well-covered.MCP Observability Explained: Monitoring AI Agent Tool Access | Obot ↗Stop Data Exfiltration with the AI & Data Security Platform (DSPM, DLP ... ↗Roo Logger — Automatically Log AI Coding Activity ↗
Enterprise AI observability commands real prices (LangSmith up to $100k+/yr, Cyberhaven funded platform), but several adjacent tools ship as open source (Roo Logger, Langfuse), creating downward price pressure on a developer-workstation agent.9 AI Observability Platforms Compared: Phoenix, LangSmith, Langfuse ... ↗Data Exfiltration via AI Agents: Attack Patterns and Gateway-Level ... ↗
AI agent security is durable: MITRE ATLAS now formally catalogs agent threats, vendors are building long-term platforms, and the Claude Code disclosure shows vendor-side opacity will keep enterprise security teams wary for years.Mitre Atlas ↗
Technically feasible via OS-level hooks (eBPF, syscall tracing, network interception) and Roo Logger proves the per-assistant pattern, but 'any AI assistant on a workstation' means juggling MCP servers, native IDE plugins, CLI wrappers, and process tracing — meaningfully harder than any single integration.Roo Logger — Automatically Log AI Coding Activity ↗MCP Observability Explained: Monitoring AI Agent Tool Access | Obot ↗