Autonomy demands
boundaries.
Five weighted pillars turn inspectable source and domain evidence into a public automated signal. A match is a lead for human investigation—not proof of a vulnerability or certification.
Security & privacy
Prompt injection, unsafe execution, secrets and network sinks · 25% of the overall signal.
SEC001Hardcoded credential
CRITICAL
A secret, such as an API key or password, appears to be written directly in the source. Anyone who can read the code, including its history, can use it.
Suggested action: Move the credential to a secret store and rotate it.
- Evidence
- Static source pattern or domain check
- Interpretation
- Heuristic finding; requires human validation.
- Maps to
- OWASP LLM02:2025 · CWE-798
SEC002Dynamic code execution
CRITICAL
The code runs text as code or as a shell command. If untrusted text such as model output or user input reaches it, an attacker can run their own commands.
Suggested action: Use an allow-listed dispatcher and argument-safe process calls.
- Evidence
- Static source pattern or domain check
- Interpretation
- Heuristic finding; requires human validation.
- Maps to
- OWASP LLM05:2025 · CWE-94 · CWE-78
SEC003Untrusted data in privileged prompt
MAJOR
Text from a request appears inside a system or developer instruction. Instructions are the part of a prompt the model trusts most, so this is a common path for prompt injection.
Suggested action: Keep untrusted data outside system or developer instructions.
- Evidence
- Static source pattern or domain check
- Interpretation
- Heuristic finding; requires human validation.
- Maps to
- OWASP LLM01:2025 · CWE-1427
SEC004User-controlled network destination
MAJOR
The code fetches a web address that a request can supply. Without checks, the agent can be steered into internal services or private network addresses.
Suggested action: Allow-list destinations and reject private and reserved IP ranges before fetching.
- Evidence
- Static source pattern or domain check
- Interpretation
- Heuristic finding; requires human validation.
- Maps to
- OWASP LLM06:2025 · CWE-918
SEC005TLS certificate checks disabled
MAJOR
The code turns off certificate checking for HTTPS connections. Anyone on the network path can then impersonate the model provider, a tool API or a data source and read or change what the agent sends and receives.
Suggested action: Keep certificate verification on; if a private CA is needed, trust that CA explicitly instead of disabling checks.
- Evidence
- Static source pattern or domain check
- Interpretation
- Heuristic finding; requires human validation.
- Maps to
- CWE-295
SEC006Unsafe deserialization
MAJOR
The code loads data with a format that can run code while it is read, such as pickle or unrestricted YAML. A poisoned file, cache entry or downloaded model then runs with the agent's permissions.
Suggested action: Use a data-only format (JSON, safetensors, yaml.safe_load) for anything that is not fully trusted.
- Evidence
- Static source pattern or domain check
- Interpretation
- Heuristic finding; requires human validation.
- Maps to
- OWASP LLM03:2025 · CWE-502
SEC007Remote model code trusted
MAJOR
The code loads a model with trust_remote_code enabled, which downloads and runs Python from the model's repository. Whoever controls that repository controls code running inside the agent.
Suggested action: Leave trust_remote_code off, or pin the model to a reviewed revision hash.
- Evidence
- Static source pattern or domain check
- Interpretation
- Heuristic finding; requires human validation.
- Maps to
- OWASP LLM03:2025 · CWE-494
SEC008Unrestricted code or shell tool
MAJOR
The agent is given a tool that runs arbitrary code or shell commands on the host, or a framework safety switch is turned off. A prompt injection in anything the agent reads can then become a command on the machine.
Suggested action: Run model-written code in an isolated sandbox (container, microVM or remote executor) with no secrets, and keep framework safety switches on.
- Evidence
- Static source pattern or domain check
- Interpretation
- Heuristic finding; requires human validation.
- Maps to
- OWASP LLM06:2025 · CWE-250
SEC009Database query built from text
MAJOR
A SQL statement is assembled by inserting values into the query text. If any of those values comes from a user or from model output, as in text-to-SQL agents, it can rewrite the query.
Suggested action: Pass values as bound parameters and allow-list identifiers such as table and column names.
- Evidence
- Static source pattern or domain check
- Interpretation
- Heuristic finding; requires human validation.
- Maps to
- OWASP LLM05:2025 · CWE-89
SEC010Model output rendered as HTML
MAJOR
Text produced by a model is inserted into a page as HTML without sanitizing it. A prompt injection can make the model emit script that runs in the user's browser.
Suggested action: Render model output as text, or pass it through an HTML sanitizer such as DOMPurify first.
- Evidence
- Static source pattern or domain check
- Interpretation
- Heuristic finding; requires human validation.
- Maps to
- OWASP LLM05:2025 · CWE-79
SEC011Untrusted data reaches code, shell or SQL
CRITICAL
Model output, tool-call arguments or an HTTP request value flows, within one function, into eval/exec, a shell command or a SQL statement. Anything that can steer the model or the request, including a prompt injection in a web page or document, can then run code or queries.
Suggested action: Do not execute model or request data. Map it to an allow-listed action, pass values as bound parameters or argument arrays, and run generated code only in an isolated sandbox.
- Evidence
- Static source pattern or domain check
- Interpretation
- Heuristic finding; requires human validation.
- Maps to
- OWASP LLM05:2025 · OWASP LLM01:2025 · CWE-94 · CWE-78 · CWE-89
MCP001Hidden instructions in an MCP tool description
CRITICAL
An MCP tool's description or docstring contains text aimed at the model rather than at people, such as <IMPORTANT> blocks, "ignore previous instructions", "do not tell the user", or paths to SSH keys and MCP configs. Clients pass tool descriptions to the model as trusted context, so this is the tool-poisoning pattern.
Suggested action: Describe what the tool does in plain language for the user; remove instructions to the model and any reference to files outside the tool's purpose.
- Evidence
- Static source pattern or domain check
- Interpretation
- Heuristic finding; requires human validation.
- Maps to
- OWASP LLM01:2025 · CWE-1427
MCP002MCP server open on all interfaces without authentication
MAJOR
An MCP server is served over HTTP on 0.0.0.0 and the file shows no authentication. Anyone who can reach the machine can list and call its tools.
Suggested action: Bind local servers to 127.0.0.1, and require authentication (for example OAuth or a bearer token) on any server reachable from a network.
- Evidence
- Static source pattern or domain check
- Interpretation
- Heuristic finding; requires human validation.
- Maps to
- OWASP LLM06:2025 · CWE-306
DEP001Dependency with a high-severity advisory
MAJOR
A dependency is pinned, in a lockfile or an exact requirement, to a version that a GitHub-reviewed advisory rates high or critical. The check uses the public OSV.dev database at scan time; whether the agent reaches the vulnerable code is not checked.
Suggested action: Upgrade to a fixed version listed in the advisory and regenerate the lockfile.
- Evidence
- Static source pattern or domain check
- Interpretation
- Heuristic finding; requires human validation.
- Maps to
- OWASP LLM03:2025 · CWE-1395
DEP002Dependency with a known advisory
MINOR
A dependency is pinned to a version with a published advisory rated moderate or low, or not yet rated. The check uses the public OSV.dev database at scan time; reachability is not checked.
Suggested action: Upgrade to a fixed version listed in the advisory and regenerate the lockfile.
- Evidence
- Static source pattern or domain check
- Interpretation
- Heuristic finding; requires human validation.
- Maps to
- OWASP LLM03:2025 · CWE-1395
Guardrails & loop control
Bounded loops, human approval and rate limits · 25% of the overall signal.
REL001LLM request without token ceiling
MAJOR
A model request does not set a maximum output length. One unusual input can produce a very long, expensive response.
Suggested action: Set max_tokens or the provider's equivalent on each model call.
- Evidence
- Static source pattern or domain check
- Interpretation
- Heuristic finding; requires human validation.
- Maps to
- OWASP LLM10:2025 · CWE-770
REL002Public LLM endpoint without rate limit
MAJOR
A public endpoint calls a model without any visible rate limit. Anyone can send many requests and run up cost or exhaust quota.
Suggested action: Apply per-user or per-IP rate limits before model calls.
- Evidence
- Static source pattern or domain check
- Interpretation
- Heuristic finding; requires human validation.
- Maps to
- OWASP LLM10:2025 · CWE-770
REL003Unbounded model loop
CRITICAL
A loop that never ends by itself contains a model call. If the agent never decides it is finished, it keeps running and keeps spending until something outside stops it.
Suggested action: Enforce a hard step limit, timeout and cancellation signal.
- Evidence
- Static source pattern or domain check
- Interpretation
- Heuristic finding; requires human validation.
- Maps to
- OWASP LLM10:2025 · CWE-835
REL004Sensitive action without human approval
MAJOR
The code can delete data, send messages or move money, and no step asks a person to approve it first. A wrong model decision becomes a real action.
Suggested action: Require explicit human approval before a destructive or external action.
- Evidence
- Static source pattern or domain check
- Interpretation
- Heuristic finding; requires human validation.
- Maps to
- OWASP LLM06:2025
Schema & tooling
Runtime validation, safe output and source provenance · 20% of the overall signal.
SCH001LLM output without runtime schema
MAJOR
Model output is used without being checked against a schema. Unexpected or malformed output can break later steps or be trusted when it should not be.
Suggested action: Validate model output with Zod, JSON Schema, Pydantic or provider structured output.
- Evidence
- Static source pattern or domain check
- Interpretation
- Heuristic finding; requires human validation.
- Maps to
- OWASP LLM05:2025 · CWE-20
SCH002Raw model text returned
MINOR
The code returns the model's raw text. Output that was never parsed or validated can pass mistakes straight to users or other systems.
Suggested action: Parse and validate the result before returning it.
- Evidence
- Static source pattern or domain check
- Interpretation
- Heuristic finding; requires human validation.
- Maps to
- OWASP LLM05:2025
SCH003RAG answer without citations
MAJOR
The code retrieves documents and generates an answer, but returns no sources with it. Readers cannot check where the answer came from.
Suggested action: Return source identifiers or citations alongside retrieved answers.
- Evidence
- Static source pattern or domain check
- Interpretation
- Heuristic finding; requires human validation.
- Maps to
- OWASP LLM09:2025
Token & cost efficiency
Model fallbacks, caching and bounded context · 15% of the overall signal.
EFF001Model call without fallback
MAJOR
A model call has no visible fallback. If the provider fails or is rate limited, the whole agent fails with it.
Suggested action: Configure a secondary model or explicit fallback path for provider failures.
- Evidence
- Static source pattern or domain check
- Interpretation
- Heuristic finding; requires human validation.
EFF002Repeated model call without cache
MINOR
An endpoint calls a model on every request with no visible cache. Repeated identical questions pay for the same answer again.
Suggested action: Cache deterministic or repeated requests with an appropriate expiry.
- Evidence
- Static source pattern or domain check
- Interpretation
- Heuristic finding; requires human validation.
- Maps to
- OWASP LLM10:2025
EFF003Unbounded conversation history
MAJOR
The full conversation history is sent to the model on every call. Cost and delay grow with each turn until the context limit is hit.
Suggested action: Trim or summarize history before sending it to a model.
- Evidence
- Static source pattern or domain check
- Interpretation
- Heuristic finding; requires human validation.
- Maps to
- OWASP LLM10:2025
EFF004Unbounded input payload
MINOR
Whole documents appear to be passed straight into a prompt. Large inputs raise cost and can crowd out the instructions.
Suggested action: Bound or compress document input before sending it to the provider.
- Evidence
- Static source pattern or domain check
- Interpretation
- Heuristic finding; requires human validation.
- Maps to
- OWASP LLM10:2025
Entity trust & provenance
DNS ownership, HTTPS and privacy disclosures · 15% of the overall signal.
TRUST001Domain ownership not verified
MAJOR
Metal Mantra could not confirm that the owner controls the stated domain. The domain may belong to someone else.
Suggested action: Verify the domain using the Metal Mantra DNS TXT challenge.
- Evidence
- Static source pattern or domain check
- Interpretation
- Heuristic finding; requires human validation.
TRUST002HTTPS not confirmed
MAJOR
The stated website could not be reached over HTTPS, so Metal Mantra could not confirm it is a secure, working site.
Suggested action: Make the official website reachable over HTTPS for verification.
- Evidence
- Static source pattern or domain check
- Interpretation
- Heuristic finding; requires human validation.
TRUST003Privacy Policy link not confirmed
MINOR
No privacy policy link was found on the stated homepage. Buyers cannot see how the vendor handles their data.
Suggested action: Expose a clear Privacy Policy link on the official homepage.
- Evidence
- Static source pattern or domain check
- Interpretation
- Heuristic finding; requires human validation.
TRUST004Terms link not confirmed
MINOR
No terms link was found on the stated homepage. Buyers cannot see the conditions that apply to using the service.
Suggested action: Expose a clear Terms link on the official homepage.
- Evidence
- Static source pattern or domain check
- Interpretation
- Heuristic finding; requires human validation.