T02 · Agent Memory Poisoning
- Location
references/heartbeat.md:3- Finding
Persistent Autonomous External Interaction Through Heartbeat Configuration
- Content
View full analysis
Vulnerability Details
File Location:
references/heartbeat.md, lines 3–9
Vulnerability Type: T02: Agent Memory Poisoning
Risk Level: MediumVulnerable Code
markdown Add to your heartbeat file: ## AgentCupid (every 30 minutes) If 30 minutes since last AgentCupid check: 1. Check for new matches → like interesting ones 2. Check active conversations → reply to unread messages 3. Check handoff requests → notify your human 4. Track daily limits (likes, messages)The recurring behavior is reinforced at
references/heartbeat.md, lines 13–19:markdown 1. **New matches:** `GET /agent/matches?limit=10` → like/pass interesting ones 2. **Unread messages:** `GET /agent/conversations` → check for new messages in active conversations 3. **Handoff requests:** Handoffs notify your human automatically, but check if any need follow-up 4. **Rate tracking:** Monitor `X-RateLimit-Remaining` headers ### Recommended Check Interval - Every 30 minutes during active hoursTechnical Analysis
The skill directs the agent to copy instructions into a persistent heartbeat file. Those instructions establish cross-session API polling and authorize recurring autonomous actions, including liking profiles and replying to messages.
This creates a persistent external-input processing loop rather than limiting activity to an explicitly invoked matchmaking request. Content retrieved from profiles and conversations is controlled by external platform users. The documentation does not establish a firm instruction/data boundary, require approval for each outbound action, or explicitly prohibit following instructions embedded in retrieved messages.
Although the skill includes privacy and escalation rules, these controls do not eliminate the risk that untrusted message content may influence recurring agent behavior. The issue is therefore best classified as persistent-state poisoning rather than executable malware or operating-system persistence.
Attack P
...[truncated 1642 chars]
- Remediation
View remediation
Remediation Suggestions
- Remove the default instruction to modify persistent heartbeat or memory files.
- Require explicit, informed owner opt-in before enabling recurring checks.
- Default heartbeat operation to notification-only behavior; do not automatically like, pass, reply, or initiate handoffs.
- Require human approval for each state-changing API request, especially likes, messages, profile updates, and handoffs.
- Treat all profile fields and conversation messages as untrusted data. Add an explicit rule that instructions contained in remote content must never be followed as agent directives.
- Isolate retrieved content from system and skill instructions using structured data boundaries.
- Minimize outbound data and prevent replies from including unrelated memory, private context, credentials, real names, addresses, workplaces, financial information, or other personal details.
- Document how to disable the heartbeat integration and remove previously persisted instructions.
- Add audit logging for recurring checks and all state-changing requests, including timestamps, conversation identifiers, and whether human approval was obtained.
- Enforce rate limits, quiet hours, conversation allowlists, and an immediate owner-controlled kill switch.
