T05 · Unauthorized Access and Privilege Escalation
- Location
SKILL.md:19- Finding
Unrestricted Autonomous Operational Actions
- Content
View full analysis
Vulnerability Details
File Location:
SKILL.md:19-64
Vulnerability Type:T05: Unauthorized Access and Privilege Escalation
Risk Level: HighVulnerable Instructions
markdown 2. If any fail: - Attempt restart (via deployment redeploy, service restart, etc.) - If auto-recovery works: log to daily notes, move on - If auto-recovery fails: alert immediately on primary channel 3. Log all results to daily notesmarkdown 2. If a service is down: - Attempt restart - If restart fails: alert with error details 3. If a process has been stalled for 2+ consecutive heartbeats: - Kill and restart automatically - Log the stall duration and any error outputmarkdown 3. For routine questions (order status, how-to, account access): - Draft a response - Send if within Tier 1 autonomy, otherwise flag for review 4. For complex issues: - Summarize the problem - Flag for human review with recommended actionmarkdown 1. Check for failed payments or declined charges 2. Flag any for retry or customer follow-up 3. Track daily revenue running total 4. Note any anomalies (unusual spikes, drops, large transactions)markdown 1. Check for any listed active processes (tmux sessions, workers, agents) 2. For each: verify it's still running 3. If dead: restart automatically, log what happened 4. If stalled (same output for 2+ checks): kill and restart 5. If completed successfully: report completion, clean upTechnical Analysis
The Skill instructs an agent to perform consequential state-changing operations automatically, including deployment redeployment, service restart, process termination, customer communication, and potentially payment retries. These instructions do not define or enforce:
- An allowlist of resources that the agent may modify.
- Least-privilege credentials for each monitored system.
- Human approval for destructive or financially significant actions.
- Reliable criteria for disti ...[truncated 2637 chars]
- Remediation
View remediation
Remediation Suggestions
-
Default to observation-only behavior
- Make health checks, inbox scans, and payment checks read-only by default.
- Require an explicit policy grant before enabling any state-changing operation.
-
Enforce resource allowlists
- Enumerate the exact services, processes, deployment targets, inboxes, and payment accounts the agent may access.
- Reject operations against resources not present in the configured allowlist.
-
Require approval for high-impact actions
- Require human confirmation before redeploying production services, terminating processes, sending customer messages, or retrying payments.
- Permit automatic restart only for specifically approved, stateless resources with tested recovery procedures.
-
Apply least-privilege credentials
- Use separate credentials for monitoring, service recovery, support communication, and payment operations.
- Do not grant write permissions to an agent whose configured task requires only status checks.
- Restrict credentials by environment, resource, operation, and time where supported.
-
Harden failure detection
- Use multiple independent signals before declaring a resource dead or stalled.
- Distinguish unchanged output from actual process failure.
- Add minimum observation windows, retry limits, cooldown periods, and circuit breakers.
-
Protect financial operations
- Require human approval for payment retries.
- Use provider-supported idempotency keys and verify transaction state before retrying.
- Define maximum retry counts and monetary thresholds.
-
Protect outbound communications
- Treat all inbox content as untrusted data rather than instructions.
- Restrict automated responses to approved templates and narrowly defined scenarios.
- Prevent messages from causing tool invocation or changes to monitoring policy.
-
Add recovery and accountability controls
- Record the triggering evidence, aut ...[truncated 447 chars]
-
