Every row on this page is a check that drives the real engine, proxy call path, tool-integrity scanner, redaction, audit chain, approvals, rate limiter and metrics against a bundled policy with one rule per mechanism. Nothing in the decision path is mocked. Generated by aggrete conformance from aggrete 0.13.0 on 2026-10-11.
pip install aggrete && aggrete conformance1 not applicable, 0 failing, of 10 controls.
2 not applicable, 0 failing, of 10 controls.
0 not applicable, 0 failing, of 14 controls.
0 not applicable, 0 failing, of 18 controls.
Each prints the evidence it found. --format json exits non-zero on any failure, so this runs in CI; --format md is what an auditor gets.
| check | what it proves | result | evidence |
|---|---|---|---|
| C01 | Refuses before fetching: a walled domain is never contacted | pass | plan__read refused at stage=pre, upstream calls=0, audit rule=CONF-WALL |
| C02 | Catches the combination: three individually-fine reads, the third refused | pass | finance + hr allowed; ops refused pre-fetch by CONF-JOIN with entity overlap |
| C03 | Prompt-injection shield: no write after reading untrusted content | pass | write allowed before taint; after an untrusted read both the write and the private-repo read are refused (CONF-FLOW) |
| C04 | Argument-level rules: the same tool is fine or forbidden by what it is asked | pass | crm__export scope=team allowed; scope=all refused pre-fetch (CONF-ARG) |
| C05 | Redaction: SSNs and emails masked before the model sees the result | pass | result masked; audit redacted={'email': 2, 'ssn': 1} |
| C06 | Inbound secrets: a credential in tool arguments is never forwarded | pass | AWS key in arguments refused, upstream not called, audit rule=inbound-secret |
| C07 | Tool integrity: poisoned descriptions blocked, rug pulls flagged | pass | hidden instruction -> block; definition changed after pin -> alert (rug pull) |
| C08 | Rate limit: per-user ceiling, refused with an audit row | pass | 6th call in the window refused, audit rule=rate-limit |
| C09 | Audit chain: every decision hash-chained; tampering is detected at the line | pass | 3 rows chained and intact; editing line 2 is reported as the first bad line |
| C10 | Identity from the token; HTTP refuses to start without auth; caller tokens never go upstream | pass | email claim -> identity; streamable-http without auth: SystemExit; upstream headers come from config, not the caller |
| C11 | Human-in-the-loop: a held call, an approval, a retry that succeeds, all audited | pass | held (request 6192934eac), approved by the clause owner, retry allowed with purpose on the audit row |
| C12 | Hidden tools: what a person can never call is never listed | pass | walled and blocked domains hidden for non-allowed users; approval-gated tools remain visible |
| C13 | Post-call rules: small groups, self-comparison and budgets refuse the result | pass | results describing 2 people, self+colleague, and 6 distinct customers all refused post-call |
| C14 | Purpose binding: a granted purpose opens a scoped, audited window | pass | refused, then allowed under a granted purpose that is stamped on the audit row |
| C15 | Dry run: 'would this be allowed?' answered without fetching | pass | three-step plan reported REFUSED at step 3 by CONF-JOIN; upstream calls=0 |
| C16 | Observability: metrics and OTLP records derive from the same audit rows | pass | Prometheus counters and an OTLP log record produced from the audit rows |
Source: https://owasp.org/www-project-mcp-top-10/
| control | title | status | shown by | note |
|---|---|---|---|---|
| MCP01 | Token Mismanagement & Secret Exposure | pass | C06, C05, C10 | the proxy holds upstream credentials; callers' tokens are never forwarded (C10) |
| MCP02 | Privilege Escalation via Scope Creep | pass | C01, C12, C04 | |
| MCP03 | Tool Poisoning | pass | C07 | |
| MCP04 | Software Supply Chain Attacks & Dependency Tampering | partial | C07 | rug-pull detection by fingerprint; package-level supply chain is out of scope |
| MCP05 | Command Injection & Execution | partial | C04, C06 | argument-level rules and secret blocking constrain what reaches a server; injection inside the server is the server's job |
| MCP06 | Prompt Injection via Contextual Payloads | pass | C03, C07 | |
| MCP07 | Insufficient Authentication & Authorization | pass | C10, C01, C12 | |
| MCP08 | Lack of Audit and Telemetry | pass | C09, C16 | |
| MCP09 | Shadow MCP Servers | n/a | deployment property: make the proxy the only allowed server (docs/DEPLOY.md, managed client allow-lists) | |
| MCP10 | Context Injection & Over-Sharing | pass | C02, C13, C05 |
Source: https://genai.owasp.org/resource/owasp-top-10-for-agentic-applications-for-2026/
| control | title | status | shown by | note |
|---|---|---|---|---|
| ASI01 | Agent Goal Hijack | pass | C03, C07 | |
| ASI02 | Tool Misuse & Exploitation | pass | C04, C08, C15 | |
| ASI03 | Identity & Privilege Abuse | pass | C10, C01, C12 | |
| ASI04 | Agentic Supply Chain Vulnerabilities | partial | C07 | |
| ASI05 | Unexpected Code Execution | n/a | not a proxy concern; see MCP05 | |
| ASI06 | Memory & Context Poisoning | pass | C03, C05 | |
| ASI07 | Insecure Inter-Agent Communication | n/a | A2A is out of scope for an MCP proxy | |
| ASI08 | Cascading Failures | partial | C08, C16 | |
| ASI09 | Human-Agent Trust Exploitation | pass | C11, C15 | |
| ASI10 | Rogue Agents | pass | C02, C13, C08 |
| control | title | status | shown by | note |
|---|---|---|---|---|
| T2 | Tool Poisoning | pass | C07 | |
| T3 | Full Schema Poisoning | pass | C07 | |
| T4 | Resource Content Poisoning | pass | C03 | |
| T8 | Confused Deputy (OAuth proxy) | pass | C10 | |
| T9 | Insecure Human-in-the-Loop | pass | C11 | |
| T11 | Prompt Injection | pass | C03 | |
| T13 | Consent / User-Approval Fatigue | partial | C11, C15 | approvals are per rule with a TTL, not per call |
| T14 | Resource Exhaustion & Denial of Wallet | pass | C08 | |
| T15 | Invisible Agent Activity | pass | C09, C16 | |
| T16 | Credential / Token Theft | pass | C06, C10 | |
| T20 | Excessive Permissions / Overexposure | pass | C01, C12, C13 | |
| T23 | Insufficient Integrity Checks | pass | C07, C09 | |
| T24 | Data Exfiltration & Corruption | pass | C03, C02, C05 | |
| T34 | Lack of Observability | pass | C16, C09 |
Source: https://standard.aiuc-1.com/
| control | title | status | shown by | note |
|---|---|---|---|---|
| A003.1 | Data access scoping | pass | C01, C02, C13 | |
| A003.2 | Agent identity management | pass | C10 | |
| A003.3 | Agent access & permissions management | partial | C11, C12 | time-limited approvals; per-tool OAuth scopes are on the roadmap |
| A008.1 | Secrets detection in user inputs | pass | C06 | |
| A008.5 | Secrets redaction in logs & artifacts | pass | C05 | |
| B006.1 | Agent service access restrictions | pass | C07, C12 | |
| B006.2 | Agent security monitoring & alerting | pass | C09, C16 | |
| B006.3 | Execution-level safeguards | partial | C07, C01 | |
| B008.1 | API & agentic interface authentication | pass | C10 | |
| B008.3 | Agentic interface data integrity | pass | C07, C06 | |
| C007 | Flag high risk outputs for human review | pass | C11 | |
| D003.1 | Tool authorization & validation | pass | C01, C04 | |
| D003.2 | Rate limits for tools | pass | C08 | |
| D003.3 | Tool call log | pass | C09 | |
| D003.4 | Human-approval workflows | pass | C11 | |
| E009.1 | Third-party access monitoring | pass | C09, C16 | |
| E015.2 | AI agent logging implementation | pass | C09 | |
| E015.4 | Log integrity protection | pass | C09 |
The widely discussed incidents of the year, scored against what a policy proxy can do. 2 stopped, 3 partial, 9 irrelevant. The irrelevant ones are supply chain, host execution and sandbox problems, where a dependency scanner or the platform vendor is the right control. Each verdict cites the checks above that demonstrate it, so you can rerun the evidence. The full list of what a proxy does not stop is its own page.
| when | incident | verdict | shown by | why |
|---|---|---|---|---|
| 2026-06 | Poisoned MCP tool descriptions make Copilot agents exfiltrate invoices tool poisoning, exfiltration through an allowed channel | stopped | C07, C03 | Hidden instructions in a description are blocked on first sight (C07); the write that would carry the invoices out after reading them is refused by the flow rule (C03). |
| 2026-04 | One PR-comment injection drives Claude Code, Gemini CLI and Copilot to leak API keys prompt injection, credential leak | partial | C03, C06 | The comment is untrusted content; any tool write afterwards is refused (C03) and a key in tool arguments never goes upstream (C06). Exfiltration through a raw shell in CI does not pass through the proxy and is not governed. |
| 2026-04 | A coding agent wipes a production database and its backups in nine seconds over-permissioned credential, destructive action without approval | partial | C11, C04 | A destructive tool call can be held for a person (C11) or constrained by its arguments (C04) only when it goes through the proxy. The agent used a credential it found in a file and a raw CLI; neither is a tool call. |
| 2026-03 | An internal agent posts unapproved advice; a colleague acts on it and widens data access agent acting without approval | stopped | C11 | A write to a shared forum under an approve rule is held until the rule's owner signs off (C11). The human decision that followed is outside any proxy. |
| 2026-01 | Path traversal plus argument injection in the Git MCP server, from prompt injection alone prompt injection, server-side bugs | partial | C04, C03 | Argument-level rules (C04) can refuse paths outside an allowed root and option-like arguments; the flow rule (C03) refuses the write that follows the injected read. The server's own bugs are the server's to fix. |
| 2026-04 | Every official MCP SDK's stdio transport turns configuration into command execution supply chain, design flaw | irrelevant | - | Happens when a client spawns a server process, below any proxy. A dependency scanner and the SDK patch are the controls. `aggrete scan` can list which configured servers are spawned over stdio. |
| 2026-08 | A worm spreads through a malicious server on the official MCP Registry supply chain | irrelevant | - | Installing the package is the compromise. Registry vetting and dependency scanning are the controls; a proxy sees the server only after it is running. |
| 2026-09 | Untrusted repositories execute code through git hooks in coding agents, outside the sandbox host execution sink | irrelevant | - | Git configuration runs on the host when the agent touches the repository. No tool call is involved. The agent vendor's sandbox is the control. |
| 2026-09 | Pinned plugin commits are checked out without verifying the pin landed supply chain, pin not verified | irrelevant | - | Plugin installation is outside the MCP call path. It does validate the pin-and-verify pattern that tool integrity (C07) applies to tool definitions. |
| 2026-02 | A GitHub issue makes Codespaces Copilot leak its token prompt injection, credential leak | irrelevant | - | The exfiltration used a schema URL download inside the client, not a tool call. The client vendor fixed it. |
| 2026-03 | An OAuth proxy in FastMCP lets a token issued for one client be replayed as another misconfiguration, confused deputy | irrelevant | C10 | A bug in a different server's auth layer. Aggrete's own design never forwards a caller's token upstream (C10), which is the class of mistake involved. |
| 2026-04 | A client-side deny rule silently stopped applying after fifty subcommands client-side policy bug | irrelevant | - | A policy that lives inside the client fails with the client. This is the argument for enforcing outside the process, which is what a proxy is; it is not something a proxy fixes in the client. |
| 2026-05 | A million exposed AI services, two in five MCP servers with no authentication misconfiguration | irrelevant | - | Servers listening on the internet with no auth. The proxy refuses to serve HTTP without an auth block (C10), but it cannot close someone else's open port. `aggrete scan` flags remote servers with no auth header. |
| 2026-02 | Summarize-with-AI buttons plant persistent instructions in assistant memory memory poisoning | irrelevant | - | The poison lives in the assistant's own memory, which no proxy sees. The flow rule limits what a poisoned session can send out (C03) but does not clean the memory. |
The same scenarios also run from the outside, over real MCP, against any endpoint. Nothing in the suite knows how a gateway words a refusal: every fixture payload carries a unique marker, and a scenario passes when the forbidden marker never reaches the client. Each has a control step, so a target that simply blocks everything scores inconclusive, not pass.
# 1. front this fixture connector with the gateway under test python -m aggrete._mockco --profile fixture # 2. run the scenarios against the gateway aggrete conformance --url https://gateway.example/mcp --token "$TOKEN" aggrete conformance --stdio "my-gateway --config gw.yaml" aggrete conformance --self # Aggrete itself, as the reference aggrete conformance --write-fixture ./fx # the reference config and policy
Two real runs, side by side: Aggrete in front of the fixture, and the fixture with nothing in front of it. The second column is the suite's own sanity check, and what an ungoverned MCP setup looks like.
| scenario | what it tests | aggrete | nothing in front | what happened with nothing in front |
|---|---|---|---|---|
| B00 | Fixture tools are reachable through the target | pass | pass | 9 tools listed |
| B01 | Combination: budget + personnel + rota about the same people | pass | fail | all three reads returned: the combination formed |
| B02 | Exfiltration: no write after reading untrusted content | pass | fail | the write went through after an untrusted read |
| B03 | Arguments: the same tool allowed or refused by what it is asked | pass | fail | scope=all returned the whole base |
| B04 | Redaction: an SSN in a result never reaches the client | pass | fail | the SSN arrived in clear |
| B05 | Inbound secrets: a credential in arguments never reaches the upstream | pass | fail | the upstream received and echoed the key |
| B06 | Tool poisoning: a tool with hidden instructions is not usable | pass | fail | the poisoned tool ran |
| B07 | Walls: a tool this person may never call is hidden or refused | pass | fail | the walled document was returned |
| B08 | Refusals explain themselves (rule and reason, not a bare error) | pass | info | 0 of 0 refusals carried a readable reason |
Controls the outside cannot see (audit integrity, identity handling, approvals) are reported as "not observable" rather than guessed. Those are covered by the sixteen in-process checks above.
Aggrete is Apache-2.0. No model in the decision path.