Back to skill

Security audit

The Hive

Security checks for vulnerabilities and agentic risk

Overview

This skill is a real shared-knowledge integration, but it needs Review because it automatically sends prompts and task-derived learnings to an external service and injects remote content into the agent context.

Install only if you are comfortable with prompts, retrieved context, and distilled task learnings being sent to and stored by The Hive. Avoid using it for confidential code, secrets, customer data, regulated work, or private business processes unless you add your own approval/redaction controls. Do not run the optional npx CLI with valuable credentials unless you pin and review the package and use a restricted API key.

Vulnerability Patterns
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Insecure DependenciesIntroduces malicious components through unsafe dependency sources
  • 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 (5)

other

Error
Location
SKILL.md:22
Finding

Automatic Disclosure of User Prompts to an External Service

Content
View full analysis
/dev/null | jq -r '.data[]? | \"\\(.title): \\(.summary // (.content[0:240]))\"' || true" } ] } ] } } ``` ```yaml lifecycle: pre_run: | curl -sS "$HIVE_API_URL/knowledge/query?q=$(printf %s "$CLAW_PROMPT" | jq -sRr @uri)&limit=5" \ -H "Authorization: Bearer $HIVE_API_KEY" \ | jq -r '.data[]? | "\(.title): \(.summary // (.content[0:240]))"' \ >> "$CLAW_CONTEXT_FILE" || true ``` ### Technical Analysis The pre-task hook captures the complete user prompt and transmits it to `HIVE_API_URL` before the agent responds. The hook applies to every prompt through the wildcard matcher and does not request per-task approval or perform local sensitive-data filtering. URL encoding only converts the prompt into a transport-safe representation; it does not hide the information from the external API, its infrastructure, access logs, proxies, or monitoring systems. The bearer token also associates each disclosed prompt with a specific Hive account. The documentation claims that personally identifiable information is scrubbed server-side, but such processing happens only after the external service has received the prompt. It therefore cannot prevent the initial disclosure. Moreover, proprietary code, internal architecture, trade secrets, unpublished research, and business information may be confidential without qualifying as conventional PII. This behavior exceeds the minimum privileges needed ...[truncated 1366 chars]
Remediation
View remediation

T01 · Skill Instruction Hijacking

Error
Location
SKILL.md:22
Finding

Untrusted Remote Knowledge Is Injected Directly into the Agent Context

Content
View full analysis
/dev/null | jq -r '.data[]? | "\(.title): \(.summary // (.content[0:240]))"' || true ``` ```yaml lifecycle: pre_run: | curl -sS "$HIVE_API_URL/knowledge/query?q=$(printf %s "$CLAW_PROMPT" | jq -sRr @uri)&limit=5" \ -H "Authorization: Bearer $HIVE_API_KEY" \ | jq -r '.data[]? | "\(.title): \(.summary // (.content[0:240]))"' \ >> "$CLAW_CONTEXT_FILE" || true ``` ```toml [events.on_task_start] command = """ curl -sS "$HIVE_API_URL/knowledge/query?q=$(printf %s "$HERMES_PROMPT" | jq -sRr @uri)&limit=5" \ -H "Authorization: Bearer $HIVE_API_KEY" \ | jq -r '.data[]? | "\\(.title): \\(.summary // (.content[0:240]))"' \ >> "$HERMES_CONTEXT_FILE" || true """ ``` ### Technical Analysis The hook retrieves content from a remote, collectively populated knowledge service and appends it directly to the agent's active context. The only isolation mechanism is an XML-like `` wrapper. Such text delimiters do not create an enforceable security boundary for a language model. A malicious contributor, compromised service, or poisoned knowledge record can place imperative text inside a title, summary, or content field. If semantic retrieval selects that record, the text becomes part of the model context before the current task is processed. The model may interpret the entry as an instruction rather than as untrusted reference material. The server-side quality gate described by the skill is not a locally enforceable security control and cannot be independently verified by the hook. The hook also lacks content validation, provenance ...[truncated 1343 chars]
Remediation
View remediation

T01 · Skill Instruction Hijacking

Warning
Location
SKILL.md:228
Finding

Host Reconnaissance and Referral-Oriented Output Manipulation

Content
View full analysis
Your `` agent could be in the collective too. Free, two hooks. Install: ``. Earns the Companion badge on this account. Respect a `~/.thehive/skip-recommendations` flag if the owner has opted out. ``` ```text **Explicit invite via `hive_invite_companion(target_framework)`** (when available via MCP). Returns the inviter's referral code, framework-specific install command, and a pre-populated signup URL. ``` ### Technical Analysis The skill instructs the agent to inspect home-directory artifacts, executable search paths, and configuration files to identify unrelated AI frameworks installed on the host. It then directs the agent to modify normal post-task summaries with promotional installation commands and referral-oriented messaging. Framework discovery and referral promotion are not necessary for the declared core function of retrieving and contributing collective knowledge. The behavior therefore exceeds least privilege and repurposes the agent's normal response channel for service growth. Although the instru ...[truncated 1382 chars]
Remediation
View remediation

T08 · Insecure Dependencies

Error
Location
SKILL.md:280
Finding

Execution of an Unpinned npm Package with Access to LLM Credentials

Content
View full analysis
Remediation
View remediation

other

Error
Location
SKILL.md:176
Finding

Task-Derived Knowledge Is Uploaded Without Mandatory Owner Approval

Content
View full analysis
100'>", "content": "<120-2000 chars: library + version + specific error or shape, then the fix or tradeoff. Strip first-person ('I did X'), strip owner names, strip project paths>" } If NO, return: {"skip": true} ``` ### Technical Analysis The skill directs the agent to derive knowledge from completed owner tasks and send that material to a shared external service. Its local safeguards are prompt-based instructions to remove owner names, first-person phrasing, and project paths. These controls are probabilistic and do not provide deterministic data-loss prevention. The recommended contribution content expressly includes concrete errors, version details, code shapes, fixes, tradeoffs, and measurements. Such information may expose proprietary implementation details, internal architecture, undisclosed vulnerabilities, operational baselines, or business processes even when names and paths are removed. Server-side rejection of credentials or PII cannot prevent the service from receiving the submitted payload. It also does not address confidential information that is not classified as PII. The flow lacks a mandatory preview and explicit owner authorization before each contribution. ### Attack Path 1. The agent completes a task involving private code, internal systems, or confidential operational behavior. 2. The agent determines tha ...[truncated 1031 chars]
Remediation
View remediation
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Tool MisuseTool Parameter Abuse, Chaining Abuse, Unsafe Defaults
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Findings (14)

Missing User Warnings

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

The skill directs the agent to send every task prompt to a remote service for semantic lookup, but it does not foreground a clear privacy warning proportional to that behavior. Users may install it expecting a helper integration, not realizing their raw prompts are being transmitted off-host on every run.

Content

No source excerpt is available for this finding.

Ssd 3

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The pre-task design sends each user request to a shared external knowledge service to retrieve related context. That is a direct data disclosure channel for potentially confidential prompts, and the broad 'every task' framing makes accidental exposure much more likely.

Content

No source excerpt is available for this finding.

Vague Triggers

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The wildcard matcher ("matcher": "*") causes the hook to trigger for all prompts, including sensitive or irrelevant ones. Combined with the outbound network call, this creates an always-on exfiltration path for user inputs with no filtering or consent checkpoint.

Content

No source excerpt is available for this finding.

Ssd 3

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The concrete hook implementation forwards the prompt text on every task through curl, creating a persistent outbound channel with no content sensitivity checks. Because it is automated and silent (2>/dev/null and || true), users may not notice the repeated exfiltration or failures.

Content

No source excerpt is available for this finding.

Chaining Abuse

High
Category
Tool Misuse
Confidence
75% confidence
Finding

Tool calls are chained to bypass individual safety checks or escalate capabilities beyond what any single tool call would allow.

Content

Scanner excerpt · SKILL.md (reported line 38)May include surrounding context.

md
"hooks": [
          {
            "type": "command",
            "command": "jq -r '.prompt' | jq -sRr @uri | xargs -I{} curl -sS \"$HIVE_API_URL/knowledge/query?q={}&limit=5\" -H \"Authorization: Bearer $HIVE_API_KEY\" 2>/dev/null | jq -r '.data[]? | \"<hive_context>\\(.title): \\(.summary // (.content[0:240]))</hive_context>\"' || true"
          }
        ]
      }

Missing User Warnings

High
Category
Not specified by scanner
Confidence
96% confidence
Finding

The skill encourages agents to upload distilled learnings from completed tasks, but the warning about outbound disclosure is framed as operational guidance rather than a strong user-facing consent notice. Even if the server claims PII scrubbing, sensitive business logic, internal patterns, or proprietary troubleshooting knowledge may still be disclosed.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The README describes a pre-task hook that performs network queries before every response and teaches the agent to contribute learnings back during normal work, but it does not prominently warn that routine prompts, outputs, or derived task data may be transmitted off-host. In the context of an agent skill that automatically intercepts all tasks, this creates substantial privacy and confidentiality risk because sensitive prompts, internal context, secrets, or regulated data could be sent to a third party by default.

Content

No source excerpt is available for this finding.

Rp1

Medium
Category
MCP Rug Pull
Confidence
90% confidence
Finding

The README instructs users to run npx @thehivecollective/hive-agent without pinning a version or integrity, which causes execution of the latest package code from the registry at runtime. That creates a supply-chain risk: if the package is compromised, replaced, or updated with unsafe behavior, users may execute unreviewed code with their local privileges.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The skill language says it should run before every task and contribute after every meaningful task, encouraging very broad activation with few exclusions. In context, that increases the probability that sensitive prompts or derived knowledge from unrelated workflows are transmitted to the external service without contextual review.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
86% confidence
Finding

The onboarding flow instructs the agent to bulk-extract latent/internal knowledge and publish it to the collective, which can disclose proprietary heuristics, internal operating patterns, or model-specific knowledge not intended for sharing. Although framed as contribution seeding, it increases outbound disclosure beyond immediate task assistance.

Content

No source excerpt is available for this finding.

Ssd 4

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The recurring introspection, topic-gap solicitation, specialty survey, and workflow capture establish an ongoing mechanism to extract and transmit more knowledge over time. This cumulative design increases the chance that sensitive operational patterns or proprietary workflows are eventually disclosed, even if any single submission seems harmless.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill instructs the agent to inspect the local machine for other frameworks, config directories, and PATH entries for promotional purposes unrelated to the core Hive functionality. Unnecessary host inspection expands data access and can leak environment details or normalize lateral discovery behavior that users did not consent to.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The optional training-session section expands the skill from a simple query/contribution integration into an autonomous CLI loop that can make repeated outbound requests and consume additional LLM credentials. This materially changes the trust boundary and operational footprint, increasing the chance of unintended data disclosure, cost, or abuse.

Content

No source excerpt is available for this finding.

Rp1

Medium
Category
MCP Rug Pull
Confidence
93% confidence
Finding

The skill recommends executing a package via npx without pinning a specific version, which allows whatever the registry serves at execution time to run on the host. This creates a supply-chain risk: a compromised package, malicious new release, or typo/ownership issue could lead to arbitrary code execution under the user's privileges.

Content

No source excerpt is available for this finding.