agent-beacon
Warn
Audited by Gen Agent Trust Hub on Oct 30, 2026
Risk Level: MEDIUMCOMMAND_EXECUTIONDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTIONPERSISTENCEPRIVILEGE_ESCALATION
Full Analysis
- [COMMAND_EXECUTION]: The skill executes a bundled Python script (
scripts/preflight.py) for integrity checks and runs thebeaconCLI for session management and configuration. - [DATA_EXFILTRATION]: The skill provides an explicit mechanism to upload session traces to a remote cloud API (
beacon share). This is a documented core feature that requires user approval for each share operation. - [INDIRECT_PROMPT_INJECTION]:
- Ingestion points: The skill reads and processes potentially untrusted content from agent session logs and history files (e.g.,
runtime.jsonlandhistory.dbdescribed inreferences/upstream.md) to facilitate trace browsing and memory review. - Boundary markers: Instructions in
SKILL.mdexplicitly warn the agent to treat all command output as data and not instructions. - Capability inventory: The skill possesses shell command execution capabilities via the
bashtool, local file modification viamemory skills install, and network egress via thebeacon sharecommand as detailed inSKILL.md. - Sanitization: The skill recommends manual preview of outgoing payloads and provides a linting tool for lenses to mitigate risks, though it notes that complete browser-based exfiltration cannot be fully prevented in lenses.
- [PERSISTENCE]: The
beacon endpoint installcommand is used to configure and start system-level background services (such as launchd or systemd) to monitor agent sessions across the host. - [PRIVILEGE_ESCALATION]: The installation of background services and management of system-wide configurations (such as those located in
/Library/Application Support/on macOS) may involve or require elevated permissions.
Audit Metadata