Back to skill

Security audit

Ai Company Ciso 2.0.0

Security checks for vulnerabilities and agentic risk

Overview

This CISO governance skill is mostly disclosed and non-executable, but its emergency approval rules contain conflicting paths for powerful crisis actions that need review before use.

Review this skill before installing in any workflow that can affect production, public communications, or agent operations. Reconcile the crisis approval rules so timeouts do not become approval, require an independent backup security approver, and narrow activation to explicit CISO governance requests.

Vulnerability Patterns
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
  • 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
Findings (1)

T05 · Unauthorized Access and Privilege Escalation

Error
Location
references/stride-assessment-crisis-channel.md:80
Finding
Crisis Workflow Permits Mandatory CISO Approval and Separation-of-Duties Bypass## Vulnerability Details **File Location**: `references/stride-assessment-crisis-channel.md:80-91` **Vulnerability Type**: Authorization bypass and privilege escalation **Risk Level**: High ### Vulnerable Documentation Segment The following is a faithful English translation of the relevant source segment: ```markdown | D-002 | CISO approval SLA timeout (>5 minutes) | High | Medium | 3.5 | Automatic escalation to direct CEO ruling + backup CISO + timeout alert | Low | D-001 additional mitigations: - Deploy the crisis direct channel on an independent high-availability cluster. - Use the CISO manual confirmation channel as a physical backup. - Degraded mode: the CISO may directly execute allowlisted operations from a terminal. ``` ### Technical Analysis The crisis protocol elsewhere requires the approval sequence `CEO → CISO approval → EXEC execution` and explicitly states that CISO approval must never be bypassed. However, the threat-assessment document defines two conflicting fallback mechanisms: 1. A CISO approval timeout can automatically escalate to a direct CEO ruling. 2. In degraded mode, the CISO can directly execute allowlisted operations from a terminal. The direct CEO ruling creates an authorization path that may omit the mandatory CISO approval. An attacker who can delay or disrupt the CISO approval channel could intentionally cause the five-minute timeout and force the workflow onto this weaker path. The terminal fallback also combines approval and execution authority in the CISO role. This defeats separation of duties between the CISO approver and the EXEC execution layer. If the CISO account or terminal is compromised, the attacker may both authorize and perform crisis operations without an independent execution control. This is a governance specification rather than executable implementation code, but an implementation following this specification would contain a security-significant fail-open authorization path. ### Attack Path 1. A ...[truncated 1546 chars]
Remediation
## Remediation Suggestions 1. Replace automatic escalation to direct CEO ruling with a fail-closed outcome. A timeout must not constitute approval. 2. Permit emergency authorization only through a separately authenticated backup CISO or an equivalent independent security approver. 3. Require cryptographically signed approval from both the initiating authority and the independent security approver before issuing an execution token. 4. Preserve separation of duties: the CISO may approve an operation but must not directly execute it. Execution should remain assigned to a distinct EXEC identity or service. 5. Enforce the crisis allowlist in a policy engine outside the agent prompt or documentation layer. 6. Bind each authorization token to the exact operation, target, parameters, requester, approver, expiration time, and incident identifier. 7. Use short-lived, single-use execution tokens and reject replayed, expired, modified, or incompletely signed requests. 8. Ensure the backup channel provides equivalent multifactor authentication, integrity protection, logging, and approval requirements rather than weaker degraded-mode controls. 9. Add automated tests confirming that CISO unavailability, denial-of-service conditions, and timeout events cannot produce an executable authorization. 10. Reconcile this threat-assessment document with the mandatory no-bypass requirements in `SKILL.md` so that all specifications define one fail-closed approval model.
Vulnerability Patterns
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • 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 (7)

Intent-Code Divergence

High
Confidence
98% confidence
Finding
The document establishes a hard security invariant that CISO approval can never be bypassed, yet later introduces an SLA-timeout path that escalates to CEO direct decision. In a crisis-control interface, this contradiction can be exploited socially or procedurally to justify execution without the required independent security gate, undermining separation of duties and enabling unauthorized privileged actions.

Intent-Code Divergence

High
Confidence
98% confidence
Finding
This section restates that execution is impossible without CISO confirmation, but that claim is incompatible with the earlier timeout rule allowing CEO direct adjudication. Such contradictory control language is dangerous because operators or automated systems may choose the weaker interpretation during emergencies, effectively creating a hidden bypass for a critical authorization control.

Intent-Code Divergence

High
Confidence
99% confidence
Finding
The final approval conditions again claim CISO approval is mandatory in all cases, while the DoS mitigation earlier permits CEO direct adjudication after timeout. Because this contradiction appears in the sign-off section, it is especially risky: it can legitimize bypass behavior in implementation, audits, or incident response and weaken assurance around high-privilege emergency operations.

Natural-Language Policy Violations

Medium
Confidence
84% confidence
Finding
The description is Chinese-only, which can implicitly force a language policy on users who did not opt into Chinese output or may misinterpret the skill's authority and constraints. In a security-governance skill, misunderstanding of approvals, prohibitions, or escalation paths can lead to operator error and unsafe decisions rather than direct code execution.

Vague Triggers

Medium
Confidence
91% confidence
Finding
The trigger list contains very broad generic terms such as '信息安全', '网络安全', '事件响应', and '安全官', which can cause the skill to activate in many unrelated contexts. In a high-authority CISO skill with network API and subagent/session capabilities, unintended invocation increases the risk of inappropriate policy steering, overbroad security instructions, or accidental disclosure into the wrong workflow.

Natural-Language Policy Violations

Low
Confidence
72% confidence
Finding
The profile fixes the communication style as Chinese-language business-oriented output, and the broader document later states technical terminology must not be used when reporting to management. While partly role-justified, it removes flexibility for user-preferred language/register and does not state that alternative language or terminology can be used when requested.

Natural-Language Policy Violations

Low
Confidence
88% confidence
Finding
This markdown file is written entirely in Chinese and presents the skill/policy content in a single fixed language. Under the stated policy, forcing a specific language without user opt-in is a natural-language locale constraint and should be documented as optional or justified.

Static analysis

No suspicious patterns detected.