autosprint

Fail

Audited by Gen Agent Trust Hub on Jul 10, 2026

Risk Level: HIGHCOMMAND_EXECUTIONPROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
  • [COMMAND_EXECUTION]: The skill provides explicit instructions to agents to retry shell commands with the sandbox disabled (dangerouslyDisableSandbox) if the codex binary fails or encounters network issues under standard containment.
  • [COMMAND_EXECUTION]: The executor workflow runs arbitrary shell commands defined in the sprint.json configuration file (checks.targeted and checks.global), which are provided as arguments to the workflow and executed without sanitization.
  • [COMMAND_EXECUTION]: The codex-lane.md reference specifies the use of the -a never flag with the Codex CLI, which deliberately suppresses approval prompts during unattended execution.
  • [PROMPT_INJECTION]: The skill implements "Worker Rules" that mandate total autonomy, explicitly forbidding agents from seeking human clarification or waiting for operator input during a run ("no agent in this run may wait on a human"), effectively removing human oversight from the execution loop.
  • [DATA_EXFILTRATION]: The skill includes built-in functionality to upload sprint dashboard artifacts and ledger data to Cloudflare R2 using the wrangler CLI. This involves the network transfer of repository-related metadata and task summaries to a well-known external service.
  • [PROMPT_INJECTION]: The skill possesses an indirect prompt injection surface as it ingests untrusted repository content—including task specs, README files, and command outputs—and interpolates this data into prompts for autonomous agents possessing command execution and PR creation capabilities.
  • Ingestion points: sprint.json, task specs, epic READMEs, and command outputs within assets/executor.workflow.js.
  • Boundary markers: The skill utilizes structured prompts but does not implement explicit delimiters or "ignore embedded instructions" warnings for the processed data.
  • Capability inventory: Spawning of sub-agents via the agent() tool, and shell execution for git, gh, codex, and arbitrary project-specific tools.
  • Sanitization: No sanitization, escaping, or validation of external content before interpolation was identified.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Jul 10, 2026, 05:11 AM
Security Audit — agent-trust-hub — autosprint