Formally-verified injection/exfiltration detector for AI agents (MCP-02).
Screen untrusted input for prompt/tool injection, exfiltration, and obfuscation before an agent consumes it. Returns a verdict (clean|suspicious|attack), probability, bits-at-risk (upper bound on adversarial capture per the Adversarial Landauer bound), matched canon patterns, and a recommended action (allow|sanitize|reject|escalate). Backed by Aristotle-verified theorems T-IB-02/T-IB-06/T-IB-01.
| Name | Required | Description | Default |
|---|---|---|---|
| input | Yes | The untrusted text/data to screen. | |
| agentId | No | Optional: for MCP-01 envelope cross-check. | |
| context | No | Optional: the agent's role/system prompt; helps calibrate. | |
| certainty | No | Operating point. Default standard. |
| Name | Required | Description |
|---|---|---|
| signals | No | |
| verdict | Yes | |
| backedBy | No | |
| bitsAtRisk | Yes | |
| probability | Yes | |
| operatingPoint | No | |
| matchedPatterns | No | |
| recommendedAction | Yes | |
| explainabilityToken | No |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations, the description fully covers behavior: it states the tool returns a verdict, probability, bits-at-risk, matched canon patterns, and recommended action. It also references Aristotle-verified theorems, adding authoritative context. No destructive behavior is implied, consistent with a screening tool.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
The description is two sentences: first states the purpose, second lists return values. It is extremely concise and front-loaded, with no wasted words.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
Given the tool has an output schema (context signal) and full schema description coverage, the description is sufficient. It explains what the tool does, when to use it, and what it returns, without needing to repeat output schema details.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema coverage is 100% with parameter descriptions, but the tool description adds value by explaining the domain (screen for injection types) and the return fields, which implicitly clarify the 'input' and 'certainty' parameters. The description enhances meaning beyond the schema alone.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The title and description clearly state the tool's purpose: screening untrusted input for prompt/tool injection, exfiltration, and obfuscation. The verb 'Screen' and resource 'untrusted input' are specific, and the description distinguishes it from potential sibling tools by detailing the types of attacks detected.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description explicitly states to use this tool 'before an agent consumes' untrusted input, providing clear context. However, it lacks explicit when-not-to-use instructions or alternatives, which is acceptable given no sibling tools exist.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Analyze an agent trace for the Gray Swan Wave 16 class: untrusted retrieved/tool output causing a tool call outside the user-declared per-turn allowlist. Returns trace counts, unauthorized tool-call evidence, canon mapping VC-AI-TOOL-0001, and claim-boundary guardrails. Backed by T-IB-25/T-IB-29/T-IB-36.
| Name | Required | Description | Default |
|---|---|---|---|
| trace | No | Single agent trace with user_prompt, allowed_tools, and events[]. | |
| traces | No | Optional batch of agent traces. | |
| targetName | No | Optional display name for the assessed target. |
| Name | Required | Description |
|---|---|---|
| mode | Yes | |
| traces | No | |
| canonId | No | |
| summary | Yes | |
| verdict | Yes | |
| probability | No | |
| theoremRefs | No | |
| claimBoundary | No | |
| reviewPriority | Yes | |
| recommendedAction | Yes | |
| explainabilityToken | No | |
| customerSystemProved | No |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations provided, the description carries the full burden. It discloses the tool's return values (trace counts, unauthorized tool-call evidence, canon mapping, claim-boundary guardrails) and backing policy references, which gives the agent a clear sense of what to expect. It does not explicitly state side-effect or permission requirements, but 'Analyze' strongly implies read-only behavior.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
The description is three sentences, front-loaded with the primary purpose, and includes return-value details and backing references. It is efficient but the backing reference (T-IB-25/T-IB-29/T-IB-36) is cryptic and may not add immediate value for an AI agent, preventing a perfect score.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
For a tool with an output schema and a well-described input schema, the description covers the core purpose, return categories, and backing context. The main gap is the absence of explicit usage guidelines relative to the sibling detect_injection, but overall it is sufficiently complete for the agent to invoke correctly.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
The input schema has 100% description coverage for all three parameters, so the baseline is 3. The tool description adds no additional semantic detail about the parameters beyond what the schema already provides, so no extra credit is warranted.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description clearly identifies the tool's function with a specific verb ('Analyze') and resource ('agent trace'), and names the exact violation class ('untrusted retrieved/tool output causing a tool call outside the user-declared per-turn allowlist'). This level of specificity distinguishes it from the sibling detect_injection.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description implies the tool is for analyzing Gray Swan Wave 16 tool-policy violations, but does not explicitly state when to use it versus detect_injection or when not to use it. There is no explicit alternative guidance or exclusion criteria, only an implied context.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Claiming proves that you control a remote MCP connector. It does not move, proxy, or interrupt the server.
Open the connector listing, choose Claim ownership, and sign in to Glama.
Complete one verification method:
GitHub identity — fastest for official registry listings. For a namespace such as io.github.alice/server, link the matching GitHub user or an account that owns the GitHub organization, then choose Claim with GitHub.
HTTP challenge — works when you can deploy a public file. Generate a token, publish the exact JSON Glama shows at /.well-known/glama.json on the same origin as the connector, then choose Check HTTP challenge.
DNS challenge — works when you control DNS but cannot change the server. Generate a token, create the exact TXT record Glama shows, wait for it to propagate, then choose Check DNS challenge.
After verification, Glama sends a confirmation email and gives you access to listing details, thumbnails, health checks, and analytics. Keep the HTTP file or DNS record in place: Glama periodically checks it and ownership remains verified while the token is discoverable.
The HTTP ownership file has this structure:
{
"$schema": "https://glama.ai/mcp/schemas/connector.json",
"claim": "glama_claim_..."
}Claim tokens are opaque, stable, and bound to the signed-in Glama account. They contain no email address or other personal information. If Glama can no longer discover a verified HTTP or DNS token, it starts a seven-day grace period before removing claim-based access. Restore the same token during that period to keep ownership verified. Never publish an email address, Glama session token, GitHub token, or connector credential as ownership proof.
If verification fails, confirm that you copied the current token exactly. The HTTP file must be public, return valid JSON with a successful HTTP response, and stay on the connector's origin. DNS changes may need more time to propagate. A claim cannot transfer to a different origin or hostname: if the connector target changes, Glama starts the grace period and the new target must be claimed separately after the previous claim is released.
For a connector linked to the official MCP Registry, registry updates continue to replace its name, description, and URL by default. After claiming, open Manage connector and enable Use Glama listing details as the source of truth if edits made on Glama should be preserved. Categories and thumbnails are always managed on Glama; registry linkage and technical connection settings continue to sync.
Control your server's listing on Glama, including description and metadata
Access analytics and receive server usage reports
Get monitoring and health status updates for your server
Feature your server to boost visibility and reach more users
To improve your MCP server's ranking:
Claim ownership of the server listing
Complete the server profile with an accurate description and thumbnail
Provide a test profile so Glama can connect to and evaluate the server
Keep tool definitions clear and complete to earn a high Tool Definition Quality Score (TDQS)
Route real usage through the Glama Gateway; more recorded successful server uses also improve the ranking
For users:
Full audit trail – every tool call is logged with inputs and outputs for compliance and debugging
Granular tool control – enable or disable individual tools per connector to limit what your AI agents can do
Centralized credential management – store and rotate API keys and OAuth tokens in one place
Change alerts – get notified when a connector changes its schema, adds or removes tools, or updates tool definitions, so nothing breaks silently
For server owners:
Proven adoption – public usage metrics on your listing show real-world traction and build trust with prospective users
Tool-level analytics – see which tools are being used most, helping you prioritize development and documentation
Direct user feedback – users can report issues and suggest improvements through the listing, giving you a channel you would not have otherwise
The connector status is unhealthy when Glama is unable to successfully connect to the server. This can happen for several reasons:
The server is experiencing an outage
The URL of the server is wrong
Credentials required to access the server are missing or invalid
If you are the owner of this MCP connector and would like to make modifications to the listing, including providing test credentials for accessing the server, please contact support@glama.ai.
No comments yet. Be the first to start the discussion!
Security firewall for AI agents — scans MCP calls for injection, secrets, and risks.
Deterministic prompt-injection detector; signed, offline-verifiable verdicts. Not an LLM.
Jailbreak-proof AI guardrails. Automated Reasoning SMT solver, not an LLM. ZK proofs included.
Decision-assurance for AI agents: an auditable action boundary + receipt before it acts.
Add one secure layer between your agents and this server.
The two tools target completely distinct aspects: one screens untrusted input for injection patterns, while the other analyzes agent traces for policy violations. There is no overlap in their purpose or output, making misselection unlikely.
Both tools follow a consistent 'detect_' + noun phrase naming convention (detect_injection, detect_trace_tool_policy). The pattern is clear and predictable, and there is no mixing of styles.
With only two tools, the set is slightly under the typical 3-15 range, but it feels appropriately scoped for a specialized security detector. Each tool covers a distinct and essential part of the domain, so the small count is justified.
The two tools cover the primary detection needs: input screening and trace analysis. Minor gaps exist (e.g., no explicit tool for sanitization or for analyzing output exfiltration), but the core injection-detection workflows are well covered with no dead ends.