Back to skill

Security audit

jep-guard-audit

Security checks for vulnerabilities and agentic risk

Overview

This is a disclosed local audit service that stores signed events in a tenant-scoped SQLite database and does not show hidden execution, exfiltration, or destructive behavior.

Install only if you intend to run a local authenticated audit API. Configure strong bearer tokens, trusted public keys, tenant quotas, and the database path carefully; run it on 127.0.0.1 unless protected by your own TLS and rate limiting. Treat exported reports and stored audit events as sensitive operational data.

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
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Memory PoisoningPersistent Context Injection, Context Window Stuffing, Memory Manipulation
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
Findings (24)

Memory Manipulation

High
Category
Memory Poisoning
Confidence
80% confidence
Finding

Skill manipulates agent memory, state, or stored context. Memory corruption can alter personality, override safety rules, or cause unpredictable behavior.

Content

Scanner excerpt · SECURITY.md (reported line 53)May include surrounding context.

md
### Upgrade from 2.0.0

Stop **all** 2.0.0 processes first: old writers do not enforce these reservations and must not share the database with 2.0.1. Make a private backup that retains every event, nonce and ingestion digest. Choose reservations for every tenant already present in the database and every tenant allowed to write. On first startup, 2.0.1 verifies existing row counts and UTF-8 body sizes against these budgets before recording the policy. Over-quota or unconfigured historical tenants cause startup to fail without deleting their records. Do not delete history or reset the database to bypass the check; use a reviewed migration that retains replay-protection evidence.

These reservations isolate retained audit data, not CPU, I/O bandwidth or total disk consumption. The byte budget counts stored event bodies, including signatures; SQLite pages, indexes, journals, backups and other programs use additional space. Provision filesystem headroom/quotas and operator monitoring separately. Enforce per-tenant request rates and timeouts at a trusted gateway for shared deployments. No automatic archival, eviction or replay-history pruning is supplied.

Tp4

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

Generating and signing audit events, performing end-to-end ingestion logic, exporting compliance recommendations, and implementing canonicalization/hashing all exceed the narrow declared purpose. The most dangerous aspect is the possibility that users think they are invoking a verifier, when they are actually invoking a system that can create or materially alter audit evidence and disclosures.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

Generating and signing audit events, performing end-to-end ingestion logic, exporting compliance recommendations, and implementing canonicalization/hashing all exceed the narrow declared purpose. The most dangerous aspect is the possibility that users think they are invoking a verifier, when they are actually invoking a system that can create or materially alter audit evidence and disclosures.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

Generating and signing audit events, performing end-to-end ingestion logic, exporting compliance recommendations, and implementing canonicalization/hashing all exceed the narrow declared purpose. The most dangerous aspect is the possibility that users think they are invoking a verifier, when they are actually invoking a system that can create or materially alter audit evidence and disclosures.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

Generating and signing audit events, performing end-to-end ingestion logic, exporting compliance recommendations, and implementing canonicalization/hashing all exceed the narrow declared purpose. The most dangerous aspect is the possibility that users think they are invoking a verifier, when they are actually invoking a system that can create or materially alter audit evidence and disclosures.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

Generating and signing audit events, performing end-to-end ingestion logic, exporting compliance recommendations, and implementing canonicalization/hashing all exceed the narrow declared purpose. The most dangerous aspect is the possibility that users think they are invoking a verifier, when they are actually invoking a system that can create or materially alter audit evidence and disclosures.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

Generating and signing audit events, performing end-to-end ingestion logic, exporting compliance recommendations, and implementing canonicalization/hashing all exceed the narrow declared purpose. The most dangerous aspect is the possibility that users think they are invoking a verifier, when they are actually invoking a system that can create or materially alter audit evidence and disclosures.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

Generating and signing audit events, performing end-to-end ingestion logic, exporting compliance recommendations, and implementing canonicalization/hashing all exceed the narrow declared purpose. The most dangerous aspect is the possibility that users think they are invoking a verifier, when they are actually invoking a system that can create or materially alter audit evidence and disclosures.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

Generating and signing audit events, performing end-to-end ingestion logic, exporting compliance recommendations, and implementing canonicalization/hashing all exceed the narrow declared purpose. The most dangerous aspect is the possibility that users think they are invoking a verifier, when they are actually invoking a system that can create or materially alter audit evidence and disclosures.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

Generating and signing audit events, performing end-to-end ingestion logic, exporting compliance recommendations, and implementing canonicalization/hashing all exceed the narrow declared purpose. The most dangerous aspect is the possibility that users think they are invoking a verifier, when they are actually invoking a system that can create or materially alter audit evidence and disclosures.

Content

No source excerpt is available for this finding.

Possible Typosquatting: 'uvicorn' resembles popular package 'gunicorn'

High
Category
Supply Chain
Confidence
70% confidence
Finding

Package name closely resembles a popular package, suggesting possible typosquatting. Attackers publish malicious packages with similar names to trick developers into installing them.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
Not specified by scanner
Confidence
96% confidence
Finding

The manifest frames the skill as an audit-verification and event-reference inspection tool. This file instead defines global storage limits, validates per-tenant quotas, enforces aggregate reservation caps, and serializes a tenant capacity policy, which is a distinct storage administration function rather than audit evidence verification or event inspection.

Content

No source excerpt is available for this finding.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
80% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · CHANGELOG.md (reported line 5)May include surrounding context.

md
## Audit behavior

SQLite stores signed input, the ingestion digest and replay identifiers in a private directory. Default path: `~/.jep-audit/events.sqlite3`; override with `JEP_AUDIT_DB`. The directory must be mode 0700 and file mode 0600. Back up the database securely. No remote persistence or automatic deletion occurs. Capacity limits are 128 events per batch and 1,000 per session. Operator-configured tenant reservations additionally bound each tenant across all sessions by event count and stored UTF-8 body bytes. The sum of all reservations cannot exceed 100,000 events or 256 MiB of event bodies. A full tenant cannot borrow another tenant's reservation. Quota checks and batch writes are one SQLite transaction; rejected batches consume no capacity. No records or replay identifiers are automatically deleted.

IDs must be canonical UUIDv4. Timestamps must include a timezone and arrive within ±300 seconds. Batches are atomic. Reading a chain is deterministic; historical timestamps are not reclassified as expired. References resolve to earlier records, parent hashes use `sha256:` plus the complete projected event digest. Client fault extensions cannot waive missing parents. Empty sessions are not valid chains. R001 checks subsequent signed verification; warnings do not establish action success.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
80% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · README.md (reported line 5)May include surrounding context.

md
## Audit behavior

SQLite stores signed input, the ingestion digest and replay identifiers in a private directory. Default path: `~/.jep-audit/events.sqlite3`; override with `JEP_AUDIT_DB`. The directory must be mode 0700 and file mode 0600. Back up the database securely. No remote persistence or automatic deletion occurs. Capacity limits are 128 events per batch and 1,000 per session. Operator-configured tenant reservations additionally bound each tenant across all sessions by event count and stored UTF-8 body bytes. The sum of all reservations cannot exceed 100,000 events or 256 MiB of event bodies. A full tenant cannot borrow another tenant's reservation. Quota checks and batch writes are one SQLite transaction; rejected batches consume no capacity. No records or replay identifiers are automatically deleted.

IDs must be canonical UUIDv4. Timestamps must include a timezone and arrive within ±300 seconds. Batches are atomic. Reading a chain is deterministic; historical timestamps are not reclassified as expired. References resolve to earlier records, parent hashes use `sha256:` plus the complete projected event digest. Client fault extensions cannot waive missing parents. Empty sessions are not valid chains. R001 checks subsequent signed verification; warnings do not establish action success.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
80% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · SECURITY.md (reported line 32)May include surrounding context.

md
## Audit behavior

SQLite stores signed input, the ingestion digest and replay identifiers in a private directory. Default path: `~/.jep-audit/events.sqlite3`; override with `JEP_AUDIT_DB`. The directory must be mode 0700 and file mode 0600. Back up the database securely. No remote persistence or automatic deletion occurs. Capacity limits are 128 events per batch and 1,000 per session. Operator-configured tenant reservations additionally bound each tenant across all sessions by event count and stored UTF-8 body bytes. The sum of all reservations cannot exceed 100,000 events or 256 MiB of event bodies. A full tenant cannot borrow another tenant's reservation. Quota checks and batch writes are one SQLite transaction; rejected batches consume no capacity. No records or replay identifiers are automatically deleted.

IDs must be canonical UUIDv4. Timestamps must include a timezone and arrive within ±300 seconds. Batches are atomic. Reading a chain is deterministic; historical timestamps are not reclassified as expired. References resolve to earlier records, parent hashes use `sha256:` plus the complete projected event digest. Client fault extensions cannot waive missing parents. Empty sessions are not valid chains. R001 checks subsequent signed verification; warnings do not establish action success.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
80% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · SKILL.md (reported line 14)May include surrounding context.

md
## Audit behavior

SQLite stores signed input, the ingestion digest and replay identifiers in a private directory. Default path: `~/.jep-audit/events.sqlite3`; override with `JEP_AUDIT_DB`. The directory must be mode 0700 and file mode 0600. Back up the database securely. No remote persistence or automatic deletion occurs. Capacity limits are 128 events per batch and 1,000 per session. Operator-configured tenant reservations additionally bound each tenant across all sessions by event count and stored UTF-8 body bytes. The sum of all reservations cannot exceed 100,000 events or 256 MiB of event bodies. A full tenant cannot borrow another tenant's reservation. Quota checks and batch writes are one SQLite transaction; rejected batches consume no capacity. No records or replay identifiers are automatically deleted.

IDs must be canonical UUIDv4. Timestamps must include a timezone and arrive within ±300 seconds. Batches are atomic. Reading a chain is deterministic; historical timestamps are not reclassified as expired. References resolve to earlier records, parent hashes use `sha256:` plus the complete projected event digest. Client fault extensions cannot waive missing parents. Empty sessions are not valid chains. R001 checks subsequent signed verification; warnings do not establish action success.

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
83% confidence
Finding

The skill metadata does not declare any tool restrictions or allowed permissions even though the described setup and the analyzer indicate capabilities involving environment access, networking, and shell/package installation. Missing explicit scope weakens least-privilege controls and can let an agent invoke broader capabilities than users or reviewers would expect.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

This README is entirely written in Chinese and provides operational instructions only in that language. Under the policy, forcing a specific language without user opt-in or a documented justification is a natural-language policy violation.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The class docstring states this is a 'One-way adapter' with 'Zero business logic. Pure field mapping + canonicalization.' However, the class also defines a reverse conversion method in L136-L180 and applies semantic policy in L67-L93 by rejecting reserved namespaces and synthesizing JAC/JEP extensions from higher-level fields. That documentation actively misstates the implemented behavior and scope.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
99% confidence
Finding

The to_jep04 docstring says the conversion 'Automatically generates nonce,' implying the adapter creates one during conversion. In reality, L49-L52 parse friendly.event_id, require canonical spelling, and reject anything not already a UUIDv4, so no nonce generation occurs. This is a direct contradiction between documented intent and actual behavior.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
86% confidence
Finding

The export endpoint generates and returns raw_data from a compliance report, which may contain sensitive audit or system information. In this file there is no confirmation prompt, user-facing log/disclosure, or comment/docstring warning that export exposes raw report contents.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The module docstring states an operator-only, non-tenant-altering scope, yet capacity_policy() builds a validated tenant quota map from provided input or the JEP_AUDIT_TENANT_QUOTAS environment variable. Adding or changing entries in that input directly changes which tenants exist in the resulting policy, which contradicts the stated claim that requests cannot create or alter tenants.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The manifest describes verification and inspection of audit evidence and tenant-scoped event references, which suggests read-oriented audit logic. This file implements environment-driven storage quota policy ingestion, which is an administrative configuration capability unrelated to the stated verification/inspection purpose.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
84% confidence
Finding

The manifest describes verifying signed audit evidence and inspecting tenant-scoped event references. While signature verification itself is in-scope, this file also reads trust anchors and bearer-token client configuration directly from process environment variables, which introduces secret/configuration access behavior beyond the stated purpose and is not documented in the manifest description.

Content

No source excerpt is available for this finding.