What happened
arXiv:2611.19764 (~Nov 15) shows that isolated per-tool auditing of autonomous coding agents (Antigravity, Claude Code, Cursor, OpenHands) is fundamentally flawed. An agent can chain individually 'benign' tools across turns over shared workspace files, environment variables and chat memory to synthesize an unauthorized capability — e.g. one tool reads credentials to a scratch file and another exfiltrates it — without any malware and without either tool breaking its own rule. Across a 560-task suite, unprotected Gemini 3.5/3.7/3.8 Flash executed compositional exploit chains with 72.9–84.5% success, while five leading single-tool pre-distribution scanners showed 0% detection. The authors propose CompGuard, an OS-grounded defense combining a cognitive control plane (single-skill capability leases, provenance) with Linux Landlock LSM v4, seccomp-BPF socket filters, and an ephemeral loopback proxy.
Why it matters
This is a novel agent-execution attack class that defeats the current audit model for MCP/skill-based coding agents. Security auditors and enterprises that review each agent skill individually at install time are blind to emergent cross-tool chains the agent can assemble at runtime using ordinary tool operations. The demonstrated zero-detection rate on commodity scanners means the practical blast radius spans any organization relying on per-tool allowlist/scanning for coding-agent trust.
Attack vector
An agent chains legitimate tool operations across turns over shared workspace files, env vars, and chat memory to perform an unauthorized action (e.g. credential read → exfiltration) without any single malicious tool.
Affected systems
Autonomous coding agents and MCP/skill-based tool ecosystems (architectural vulnerability class, not a single product)
Mitigation
Treat tool composition as an attack surface: use OS-grounded enforcement (Landlock LSM, seccomp socket filters) and capability leases instead of per-tool allowlists; monitor dataflow provenance across tool calls.