Back to skill

Security audit

Ship Loop

Security checks for vulnerabilities and agentic risk

Overview

This skill appears to be a real build-and-deploy automation tool, but it can let an autonomous agent change, commit, push, and sometimes discard repository work without enough safeguards.

Install only for trusted repositories and run from a clean branch after committing or stashing all work. Review SHIPLOOP.yml as executable code, avoid permission-bypass agent modes unless necessary, disable meta or optimization learnings for untrusted projects, and update the docs dependencies before building docs.

Vulnerability Patterns
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
Findings (2)

T09 · Insecure Skill Coding Practices

Error
Location
shiploop/git_ops.py:160
Finding

Automatic Meta-Loop Cleanup Can Permanently Delete Uncommitted Work

Content
View full analysis
None: await run_git(["checkout", "--", "."], cwd, check=False) await run_git(["clean", "-fd"], cwd, check=False) ``` ```python # shiploop/loops/meta.py:44-56 async def run_meta_loop( config: ShipLoopConfig, segment_name: str, segment_prompt: str, all_errors: list[str], reporter: Reporter, budget: BudgetTracker, learnings: LearningsEngine, ) -> MetaResult: repo = Path(config.repo) meta_config = config.meta num_experiments = meta_config.experiments if not meta_config.enabled: reporter._print(" Meta loop disabled — segment failed") return MetaResult(success=False) reporter.meta_start() await discard_changes(repo) ``` ```bash # scripts/meta-experiment.sh:64-68 # Create experiment branch # -------------------------------------------------------------------------- # Clean working tree (caller already discarded failed changes, but be safe) git checkout -- . 2>/dev/null || true git clean -fd 2>/dev/null || true ``` ### Technical Analysis `git checkout -- .` discards modifications to tracked files, while `git clean -fd` permanently deletes all untracked, non-ignored files and directories beneath the repository. The cleanup is performed directly against the operator-configured repository before meta experiments are isolated in worktrees. No clean-tree precondition, interactive confirmation, backup, stash, path restriction, or dry-run validation protects unrelated user work. The commands also use `check=False` or suppress errors, making it difficult for the caller to distinguish complete c ...[truncated 1753 chars]
Remediation
View remediation

T02 · Agent Memory Poisoning

Error
Location
shiploop/loops/optimize.py:145
Finding

Untrusted Build Output Can Poison Reusable Agent Instructions

Content
View full analysis
str: truncated_error = preflight_error[:500] if len(preflight_error) > 500 else preflight_error truncated_diff = repair_diff[:500] if len(repair_diff) > 500 else repair_diff return f"""## OPTIMIZATION ANALYSIS for segment: {segment_name} The segment shipped successfully, but only after repair. Your job is to figure out what modifications to ...[truncated 6374 chars]
Remediation
View remediation
Vulnerability Patterns
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Findings (93)

Known Vulnerable Dependency: astro==6.0.8 — 10 advisory(ies): GHSA-26w7-cxv4-gfx2 (Astro: Remote code execution through AVIF image optimization); CVE-2026-54299 (Astro: Host header SSRF in prerendered error page fetch); CVE-2026-84376 (Astro: Authorization bypass from missing path-segment boundary check when stripp) +7 more

Critical
Category
Supply Chain
Confidence
97% confidence
Finding

The lockfile pins astro 6.0.8, and the cited advisories include severe issues such as remote code execution and SSRF. In this skill’s context, Astro is part of a docs/build pipeline, so vulnerable behavior can be reachable during local builds, preview/dev servers, content processing, or deployment workflows, making this a real supply-chain risk rather than a theoretical issue.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The declared description describes an orchestration skill for managing a multi-step software delivery pipeline across multiple feature segments. The supplied code does not implement any pipeline logic, task sequencing, deployment, verification, notification, or agent coordination. Instead, it is a simple Astro/Starlight configuration file that registers a documentation content collection with a loader and schema. This is a materially different primary purpose, so the description does not accurately represent the code.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The declared description says this skill orchestrates a multi-step delivery pipeline across feature segments, including building, shipping, verifying, notifying, and preventing handoff drops. The supplied code does none of that. It is a shell script for recording and retrieving structured 'learnings' entries in a YAML file, with stop-word filtering and keyword matching to surface relevant past failures. This is a materially different primary purpose, and the persistence/readback behavior is undeclared. Therefore the description does not accurately represent the code.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

The declared description promises an orchestrator for a multi-step, multi-segment pipeline with build, ship, verify, and notify handoffs across sequential feature tasks. The actual script is much narrower and materially different: it is a recovery/experimentation utility for a failed segment. It reads config, creates an experiment branch from a base commit, runs an agent command with a prompt, invokes preflight validation, and prints structured pass/fail output. There is no evidence of commit orchestration, deployment/shipping, verification beyond preflight, notification, or progression across multiple segments. This is a substantive description-behavior mismatch, not just an implementation-detail difference.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The declared description describes an orchestrator for sequential feature delivery segments, including building, shipping/deploying, verification, notification, and preventing handoff drops across segments. The supplied code does none of that. It is a standalone preflight script whose sole function is to run configured build, lint, and test commands before commit and fail fast if checks do not pass. It neither performs deployment/shipping, notification, segment chaining, nor multi-step agent task coordination. This is a materially different primary purpose, so it is a clear mismatch.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
94% confidence
Finding

The declared description presents a broad orchestration skill for sequential feature delivery, including build, ship, verify, and notify stages for each segment. The supplied code chunk is much narrower: it is a shell script specifically for repairing a failed segment by constructing a prompt from an error log, git diff, and untracked files, then calling an external agent command. While this could be a supporting component within such a pipeline, the code chunk itself does not implement the declared end-to-end pipeline behavior and instead provides an undeclared repair capability. Therefore the description does not accurately represent this code chunk's actual behavior.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
94% confidence
Finding

The description partially overlaps with the code in that this script is part of a ship/verify style workflow for segmented work. However, the declared purpose says it runs a chained build→ship→verify→notify pipeline for multi-segment feature work, emphasizing sequential feature handoffs. The actual code is specifically a single-segment orchestrator and does much more than a standard pipeline: it executes an agent, retries via repair loops, performs meta-analysis, creates and merges experiment branches, records learnings, and modifies repository state. There is also no visible notify step in this chunk, and multi-segment orchestration is not implemented here. These are material behavior differences, not just supporting details, so this is a mismatch.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
94% confidence
Finding

The declared description presents a broader orchestration skill for multi-segment feature delivery, covering build, ship, verify, and notify across multiple sequential feature tasks. The supplied code chunk is a single shell script focused narrowly on shipping one segment via git operations plus optional verification. While its commit/push/tag/verify behavior partially overlaps the 'ship→verify' part of the description, it lacks major declared functionality: there is no build stage, no notification stage, no loop/chaining over multiple segments, and no execution of coding-agent tasks. Therefore the code does not accurately represent the declared end-to-end pipeline purpose.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The declared description promises a full multi-stage, multi-segment pipeline orchestration workflow (build, ship, verify, notify, sequential feature handling, and handoff protection). The supplied code chunk is a single shell script focused narrowly on verifying a deployment by polling URLs and optional health/marker/header signals from a configured site. While 'verify' is one component mentioned in the description, the code does not implement the broader pipeline orchestration, segment sequencing, committing, deploying, or notifying behavior. This is therefore a material description-versus-behavior mismatch.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The declared description promises a substantial orchestration capability for sequential feature delivery, but the supplied code chunk contains only a package version declaration. There is no evidence of pipeline execution, agent-task chaining, deployment, verification, notifications, or handoff management. This is a clear description-to-behavior mismatch.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

The description claims a multi-step orchestration skill for sequential feature delivery with commit, deploy, verify, and notify handoffs. The supplied code does not implement that pipeline. Instead, it is a lower-level utility for launching an external agent subprocess, passing it a prompt over stdin, collecting output, handling timeouts, saving logs, and accounting for usage/cost. Those behaviors are materially narrower and different from the declared end-to-end pipeline purpose, and the subprocess execution/logging capabilities are undeclared. Therefore this is a description-behavior mismatch.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

The declared description is about orchestrating a multi-stage build/ship/verify/notify workflow for sequential feature delivery. The supplied code does not implement pipeline orchestration, feature sequencing, commits, deployment, verification, notification, or handoff control. Instead, it provides a budget tracker focused on LLM token usage, estimated costs, storage of usage records, and budget checks. While such budgeting could support a larger pipeline system, this chunk’s actual primary purpose is materially different from the declared skill behavior, so this is a clear mismatch.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
94% confidence
Finding

The supplied code chunk is narrowly scoped to deployment verification. It dynamically loads a provider-specific verifier and calls its verify method with deployment details. There is no evidence here of orchestrating a chained build→ship→verify→notify workflow, sequencing multiple feature segments, managing coding agent tasks, committing changes, deploying them, or notifying stakeholders. While verification is one component mentioned in the description, the code’s primary purpose is substantially narrower than the declared end-to-end pipeline behavior, so this is a description-to-behavior mismatch.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
94% confidence
Finding

The declared description presents a higher-level orchestration skill for a multi-step build→ship→verify→notify workflow across feature segments. This code chunk instead implements Git helper functions and a file-pattern security scan. While Git operations could support such a pipeline, the actual code shown does not perform the advertised pipeline stages like build, deployment, verification, or notification, nor does it coordinate multi-segment progression. It also includes repository-manipulation and secret-blocking behavior that is not explicitly declared. This is a material description-versus-behavior mismatch rather than a mere supporting-detail difference.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
93% confidence
Finding

The code does not implement a build/ship/verify/notify pipeline or multi-segment task orchestration. Its primary purpose is to capture and retrieve operational learnings from prior runs, including storage, scoring, search, and prompt formatting. While such functionality could support a larger segmented workflow, this chunk itself is a memory/lessons subsystem, which is materially different from the declared orchestration purpose.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
86% confidence
Finding

The declared description says this skill orchestrates sequential multi-feature work with commit/deploy/verify/notify handoffs between segments. The supplied code instead handles a fallback 'meta loop' for one segment after repeated failures. It asks an agent to analyze failures, produce multiple alternative implementation prompts, runs each in isolated worktrees/branches, preflights them, chooses a winner, and merges it. While this is related to software task orchestration, its primary purpose is materially different from the declared pipeline. Key declared stages—shipping/deployment, verification of deployed work, notification, and sequencing across multiple segments—are absent from this code chunk.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

The declared description is about orchestrating a multi-segment delivery pipeline with build, ship, verify, and notify handoffs. This code does something materially different: it analyzes a failed-then-repaired segment, generates multiple prompt variations, runs those variations in isolated worktrees, checks them with preflight, selects a winner based on passing status and diff size, and stores the resulting optimization learning. While it may be part of a larger system, this chunk’s primary purpose is prompt optimization experimentation, not sequencing feature segments through commit/deploy/verify/notify. That is a substantial description-behavior mismatch.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The declared description is about orchestrating an operational multi-step delivery pipeline across feature segments. The supplied code does not build, ship, verify, notify, commit, deploy, or coordinate sequential coding tasks. Its primary purpose is retrospective analysis of past runs and generation of recommendations, plus optional persistence of auto-generated learning records. This is a materially different function and includes undeclared database analytics and write-back behavior, so the description does not accurately represent the code.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

The declared description says this skill coordinates an end-to-end multi-feature pipeline where each segment is implemented, committed, deployed, verified, and notified before proceeding. The supplied code does not do that. It is narrowly focused on repair: given a failed preflight result, it repeatedly invokes an agent to fix issues, reruns preflight checks, detects repeated identical errors, enforces budget limits, and records learnings or decision gaps. There is no evidence of segment chaining, build/ship orchestration beyond rerunning preflight, commits, deployments, or notifications. The code’s primary purpose is materially different from the declared pipeline orchestration, so this is a clear description-behavior mismatch.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The declared description promises an end-to-end chained build→ship→verify→notify workflow for multi-segment feature implementation, including commits, deployment, verification, and notifications between segments. The supplied code only performs a local preflight sequence of build, lint, and test commands, capturing output and halting on failure or timeout. There is no logic for shipping/deploying, notifying, coordinating multiple feature segments, managing coding-agent tasks, or preventing handoff drops. This is a materially different and much narrower primary purpose than declared, so it should be flagged as a mismatch.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
90% confidence
Finding

The supplied code is a narrow abstraction layer for deployment verification: it declares a Pydantic result model and an abstract verifier interface. It does not implement a full pipeline, orchestrate multiple feature segments, commit work, deploy artifacts, or send notifications. While verification is one component mentioned in the description, the primary declared purpose is a broader end-to-end sequential workflow, which is not represented by this code chunk.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The declared description presents an end-to-end orchestration skill for sequential feature delivery, covering build, deploy, verify, and notify across multiple segments. The supplied code only defines a custom verifier provider that executes a configured shell command to validate a deployment target, passing limited environment variables and returning success/failure details. This is a materially different and much narrower behavior than the declared pipeline orchestration purpose, so the description does not accurately represent the code.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

The declared description presents a comprehensive orchestration skill for sequential multi-feature delivery, including build, ship, verify, and notify stages with handoff protection. The supplied code chunk does not implement that pipeline. It is narrowly focused on one subcomponent: verifying a Vercel deployment by polling URLs, checking status codes, headers, and optional content markers until success or timeout. While verification is a supporting part of the declared workflow, this chunk’s actual primary purpose is materially narrower and different from the declared end-to-end orchestration behavior. Therefore, the description does not accurately represent what this specific code chunk actually does.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

The declared description promises a high-level workflow skill that coordinates sequential feature implementation steps with commit, deploy, verify, and notify stages. The actual code shown does not implement or test such a pipeline. Instead, it tests lower-level agent helper behavior: executing shell commands, passing prompts via stdin, handling nonzero exits and timeouts, persisting outputs to log files, and recording usage/cost metrics. These behaviors are materially different from the claimed purpose, so this is a clear description-behavior mismatch.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The declared description is about orchestrating a multi-stage feature delivery pipeline (build, ship, verify, notify) across sequential feature segments. The supplied code does not implement or exercise such orchestration behavior. Instead, it contains Python tests for budgeting-related components: recording usage, checking budget thresholds, persisting spend data, summarizing costs, parsing token usage strings, and estimating token/cost values. These are materially different capabilities and indicate a different primary purpose than the declared pipeline coordination skill.

Content

No source excerpt is available for this finding.