slack-notifications

Warn

Audited by Gen Agent Trust Hub on Nov 13, 2026

Risk Level: MEDIUMMETADATA_POISONINGPROMPT_INJECTIONINDIRECT_PROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
  • [METADATA_POISONING]: The skill's YAML frontmatter contains self-asserted and unverifiable claims including 'source: builtin' and 'trust_score: 100'. These assertions deceptively signal official verification and a high level of security that are not independently validated, potentially misleading agents or users about the skill's provenance.\n- [PROMPT_INJECTION]: The 'Memory Protocol' in SKILL.md and the instructions in commands/slack-notifications.md use forceful language ('MANDATORY', 'ASSUME INTERRUPTION', 'follow it exactly as presented') to override the agent's default operational behavior. These patterns are designed to compel the agent to follow a non-standard workflow for state persistence in local files and to strictly adhere to the skill's instructions without deviation.\n- [INDIRECT_PROMPT_INJECTION]: The skill provides tools like 'channel-history' and 'get-reactions' that ingest untrusted data from Slack, exposing the agent to potential manipulation via external content.\n
  • Ingestion points: External message content and reactions retrieved from Slack channels via API calls defined in SKILL.md.\n
  • Boundary markers: Absent. The skill instructions do not provide delimiters or warnings to help the agent distinguish between its system instructions and data contained within Slack messages.\n
  • Capability inventory: The agent possesses powerful capabilities including file uploads ('upload-file'), message deletion ('delete-message'), and access to the system shell ('Bash'), which could be abused if the agent obeys instructions embedded in Slack messages.\n
  • Sanitization: Absent. The hooks/pre-execute.cjs script contains no validation logic for incoming data, and the SKILL.md provides no guidance for filtering or escaping ingested content.\n- [DATA_EXFILTRATION]: The 'upload-file' tool facilitates the transfer of local files to Slack. While intended for legitimate messaging purposes, this capability provides a technical vector for exfiltrating sensitive local data (such as credentials or configuration files) if the agent is manipulated via prompt injection. This risk is noted as a capability surface, partially mitigated by the requirement for user confirmation.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Nov 13, 2026, 03:11 AM