Back to skill

Security audit

Agentsop Code Execution Decision

Security checks for vulnerabilities and agentic risk

Overview

This is a documentation-only skill that gives a mostly well-scoped rubric for when agents should use sandboxed code execution for deterministic computation.

Before installing, note that this skill encourages code execution for exact computations, so pair it with a real sandbox and avoid exposing unsandboxed REPL tools to untrusted input. The mixed-language headings may be confusing but do not change the security posture.

Vulnerability Patterns
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • 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
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Findings (10)

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

This markdown skill includes section headings in Chinese such as '何时激活' while the rest of the operational guidance is in English, and there is no statement that the user opted into bilingual output or that the skill is region-specific. That can violate a language/locale policy requiring skills to respect user language preferences rather than implicitly imposing a locale.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The heading '核心心智模型' introduces a locale-specific presentation choice in an otherwise English skill, but the file does not justify this as a region-specific skill or offer any language selection. Repeated mixed-locale headings can create inconsistent user experience and conflict with organizational language policy.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The section title 'SOP 工作流' is a natural-language locale choice embedded in the skill instructions, but the manifest and body do not explain why Chinese is required or optional. Under the policy, forcing or assuming a language without opt-in is a violation.

Content

No source excerpt is available for this finding.

Unbounded Resource Access

Medium
Category
Excessive Agency
Confidence
75% confidence
Finding

Skill allows unbounded resource consumption (API calls, storage, compute). Without rate limits or quotas, a compromised or misbehaving agent can cause denial-of-service or cost overruns.

Content

Scanner excerpt · SKILL.md (reported line 149)May include surrounding context.

md
1. **Sandbox choice** (see §4 op #3, cross-link [[agentsop-output-format-by-model]] for the wire shape): DSPy `ProgramOfThought` uses a Python interpreter (Deno/PythonInterpreter sandbox in recent versions); OpenAI Code Interpreter runs in a managed container; Anthropic code-execution tool runs in a sandboxed VM; LangChain `PythonREPLTool` runs **in-process and is unsandboxed** — treat as untrusted-input-hostile.
2. **Run** the code in the sandbox; capture stdout / return value / traceback.
3. **Feed the result back into the LM** — this is the step under-coders forget. The code produces a *value*; the LM must consume it to narrate, format, or chain into the next step. PoT is "code computes, LM contextualizes," not "code replaces the LM."
4. **Error → retry** (cross-link [[agentsop-test-fix-loop]]): on traceback, feed the error back to the LM, regenerate the code, re-run. Bound the retries (DSPy `ProgramOfThought` defaults to `max_iters` ≈ 3). After the bound, fall back to prose reasoning or surface the failure — do not loop forever.

**Exit criterion:** the step produces either (a) a code-derived value the LM has consumed, or
(b) a prose judgment, with the gate decision recorded so a reviewer can audit *why* code was or

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The heading '操作模型' is another locale-specific instruction label in a skill otherwise written for a broad English-speaking audience. Because the file provides no user choice or regional justification, it presents a natural-language policy concern.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
87% confidence
Finding

The heading '困境决策案例' continues the pattern of mixed-language presentation in core instructional content. Without an explicit bilingual policy or locale scope, this can be treated as an imposed language choice.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
87% confidence
Finding

The heading '反模式与边界' is a language-specific instruction element, but the file does not indicate that multilingual headings are expected or user-selectable. This creates a policy issue around implicit locale enforcement.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
86% confidence
Finding

Using '跨框架对照' in an otherwise English operational document introduces a locale-specific presentation without explaining the reason or giving the user a choice. This can violate language/locale policy expectations for consistent, opt-in language behavior.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The trigger "any step about to be reasoned" is extremely broad and can cause this overlay to activate on nearly every reasoning step, not just those involving deterministic computation. In an agent system, overly broad activation can create unsafe or unintended routing into code-execution pathways, increasing the chance of unnecessary tool use and expanding the attack surface when later operations include runtime selection such as unsandboxed REPL options.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
84% confidence
Finding

The heading '引用源' is another example of locale-specific language appearing in core documentation without an explicit multilingual policy. Although low impact, it still reflects an undocumented language choice.

Content

No source excerpt is available for this finding.