Back to skill

Security audit

Hippocampus

Security checks for vulnerabilities and agentic risk

Overview

This skill is review-worthy because it broadly gathers and persists operational context across workspaces and also changes learning files beyond the headline HIPPOCAMPUS update.

Install only where agents are intended to read the listed OpenClaw workspaces and where persisting summaries of CRM, family conversation-derived data, journals, cron logs, and peer context is acceptable. Review the .learnings decay behavior first, because it can change and archive files outside HIPPOCAMPUS.md.

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

Warning
Location
SKILL.md:17
Finding

Cross-Workspace Collection and Persistence of Sensitive Operational Data

Content
View full analysis
datetime('now', '-24 hours') AND bu IN ('{bu}', 'portfolio') ORDER BY ts DESC" ``` Replace `{bu}` with your BU. Events override memory files on contradiction. Bobina: skip — sandboxed. Read from memory files and reports where results are already captured; do not query external APIs (Supabase, PostHog, GSC, Notion, Hashnode, Agent Commune). ``` `SKILL.md:37-43`: ```markdown **Peer Compression:** Information from outside your own workspace (peer HIPPOCAMPUS files, BU captures, cross-workspace reads) compresses to one line per item in your output. If you can't say it in one line, you're absorbing detail that belongs to the owning agent. Your hippocampus carries your domain at depth; peer state is directional context only. **Source Priority:** 1. **Own workspace** (memory, artifacts, data) — full fidelity. 2. **Activity feed** (events.db) — full fidelity for your BU events matching your Track list. 3. **Peer HIPPOCAMPUS** (where listed in Sources) — read for state awareness, compress to 1 line per peer. 4. **BU captures** (where listed) — filter entries to tags matching your domain. See per-agent capture filters in registry. ``` `SKILL.md:58-62`: ```markdown `memory/YYYY-MM-DD*.md` — filter by filename, don't read all files. ### Step 4: Read domain-specific sources Per registry. Rules: 14d window by filename, cross-workspace is read-only, CRM hubs selective (Bobo only for Open Thread entities), cron logs last 20 lines. Cross-workspace reads are context, not content. Extract only what maps to your T ...[truncated 5409 chars]
Remediation
View remediation
Vulnerability Patterns
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Rogue AgentSelf-Modification, Session Persistence
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
Findings (15)

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The declared purpose is a 14-day HIPPOCAMPUS sync, but the documented behavior also performs unrelated maintenance actions such as decay processing, archival, and modifications to .learnings files. This mismatch is dangerous because operators may authorize the skill expecting a narrow update while it silently performs broader state-changing actions across the workspace.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

High
Category
Not specified by scanner
Confidence
88% confidence
Finding

The file asserts that bobina is 'sandboxed' to relative paths only, while the same shared registry normalizes absolute paths, home-directory paths, and parent-directory traversal for other agents. That mismatch is dangerous because it can create a false sense of isolation in a component selected from the same registry, and increases the chance that implementation code applies broader path semantics than intended or that an attacker pivots through inconsistent path handling.

Content

No source excerpt is available for this finding.

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
94% confidence
Finding

The skill directs file reads across the workspace and from shared locations, but it does not declare an explicit tool scope or permissions boundary. That makes its effective access broader and less auditable than the manifest suggests, increasing the risk of unintended data exposure or overreach when invoked universally.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The manifest frames the skill as a daily HIPPOCAMPUS update, yet the process includes a separate maintenance workflow that changes other files and archives data. Hidden secondary operations undermine user consent and make review harder, especially in a universal skill likely to run often.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The skill is described as universal for all agents and auto-detects identity, which creates a broad activation surface. Overly broad invocation increases the chance of accidental execution in the wrong context, leading to unintended reads, writes, or persistence updates across many agent environments.

Content

No source excerpt is available for this finding.

Session Persistence

Medium
Category
Rogue Agent
Confidence
88% confidence
Finding

The instruction to write a minimal HIPPOCAMPUS from the agent's own memory only encourages persistence of unsourced session-derived information into workspace files. That can leak transient context, inaccuracies, or sensitive prior-session data into durable storage without verification.

Content

Scanner excerpt · SKILL.md (reported line 11)May include surrounding context.

md
Sachee reads these files to make decisions. A bad hippocampus wastes his time; a good one surfaces what needs attention and why.

Daily incremental update of HIPPOCAMPUS.md for all 13 agents. Detects agent ID from `IDENTITY.md` `**ID:**` field, looks up domain config below, runs universal process with domain-specific sources and framing. If no ID match, write minimal HIPPOCAMPUS from own memory only.

---

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The instructions authorize writes to HIPPOCAMPUS.md, memory files, and archival locations without an explicit up-front warning that workspace files will be modified. This can cause silent state changes and persistence side effects that a user did not knowingly approve.

Content

No source excerpt is available for this finding.

Session Persistence

Medium
Category
Rogue Agent
Confidence
90% confidence
Finding

The skill explicitly writes HIPPOCAMPUS.md as a durable artifact, which creates session persistence by design. In context this is expected functionality, but it is still security-relevant because the file may accumulate sensitive operational context and influence future decisions if poisoned or populated with incorrect data.

Content

Scanner excerpt · SKILL.md (reported line 88)May include surrounding context.

md
### Step 10: Write, verify, deliver
1. Read `references/examples.md` for before/after examples of good hippocampus entries.
2. Write HIPPOCAMPUS.md. Timestamp: `> Last updated: YYYY-MM-DD HH:MM AEST`
3. Verify size vs target. Over? Compress.
4. Verify accuracy: verify every date against source files — a wrong date is worse than no date. Threads real, commitments sourced, domain data matches files.
5. **Domain enforcement check:** Review every item in Top of Mind, Open Threads, Recent Sessions, and Commitments against your registry's Track/Exclude lists. Delete any item that falls under Exclude — even if it appeared in a source file tagged with your agent name. Common violations: content agents absorbing deal events, pipeline agents absorbing infrastructure changes, JDN agents absorbing advisory data. If in doubt, ask: "Is this my domain, or am I echoing another agent's work?"

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The skill instructs execution of a local Python script even though its stated purpose is document synchronization. Executing local code expands the attack surface because reviewers cannot infer the script's full behavior from this file alone, and the script may read, modify, or exfiltrate additional data.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
84% confidence
Finding

The comment says hippocampus-sync reads this file at Step 2 to detect the current agent's domain config, but it does not define strong scoping or selection rules. In a universal skill serving many agents, ambiguous agent selection can cause the wrong profile, sources, exclusions, or access paths to be used, leading to unintended data exposure or cross-domain context mixing.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The registry includes an explicit shell command (cd ... && git log --oneline --since="14 days ago") and git branch access for the fernando agent, which expands this skill from a markdown sync into command-execution guidance. In a universal skill that reads per-agent config from a shared registry, documenting executable commands creates a path for unintended command use, prompt-influenced execution, or future unsafe expansion beyond simple file reads.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The script writes to multiple files under .learnings and to monthly archive files, while the skill metadata says it performs a daily incremental update of HIPPOCAMPUS.md. This mismatch can violate operator expectations and trust boundaries: an agent or user enabling a seemingly narrow context-sync skill may unknowingly grant broader write behavior that alters or deletes historical records in adjacent files.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Low
Category
Not specified by scanner
Confidence
66% confidence
Finding

Section 4 states that cross-workspace reads are read-only context, reinforcing a limited read posture for external sources. Later, the process instructs appending to memory/YYYY-MM-DD.md, which introduces an additional write side effect not reflected in that read-only framing and can mislead readers about the skill's actual mutation behavior.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
36% confidence
Finding

The phrase 'Relationships (per-person: warmth, last/next action, trigger)' is natural-language guidance embedded in the file and appears to prescribe a fixed interpersonal labeling scheme. While SQP-3 is narrowly scoped, this is the only plausible policy-adjacent language in the file and may be seen as an organizationally questionable communication framing.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Low
Category
Not specified by scanner
Confidence
75% confidence
Finding

The docstring describes archiving stale entries by month, which implies a straightforward archival action, but the implementation aggregates archived blocks from all target files into a single per-month file under .learnings/archive. This loses source-file separation and is a meaningful divergence from the documented behavior.

Content

No source excerpt is available for this finding.