# OWASP and NIST mappings

These tables say where signafAIde's records and tool-call checks are evidence for an item in two public lists: the OWASP Top 10 for Agentic Applications 2026 and the audit controls of NIST SP 800-53 Rev. 5. They describe what is built today. They are not a statement that a system is compliant with, certified against or graded under either list: a record attests integrity and timing, never that a decision was right, and a control is met by a whole system and its organisation, not by one tool.

Every row has a **Not covered** column. Read it before you rely on the row.

Related pages: [Tool-call checks](/docs/checks), [Rules](/docs/rules), [Rulings](/docs/rulings), [Tool fingerprints](/docs/fingerprints), [Tool coverage](/docs/coverage), [Verify](/docs/verify). For the auditor's view, see [Auditors and compliance](/auditors).

## Words used in the tables

- **Recorded call**: a call to a tool that goes through an SDK wrapper, a framework adapter or the MCP gateway. Each one writes a `decision` record before the tool runs and an `outcome` record after it, signed by the agent and co-signed by the witness. Anything the application does around them is not recorded ([Tool coverage](/docs/coverage)).
- **Tamper-evident, not tamper-proof**: the chain, the signatures and the public timestamps make an edited, removed or reordered record show when `signafaide verify` runs. They do not stop anyone from editing a copy, or from deleting one.
- **Checks** run on the call before the tool, and while Calibrating they never block. Central rules and rulings block only on tools you protected ([Tool-call checks](/docs/checks)).
- By default a record holds a salted hash of its content, not the content ([Concepts](/docs/concepts)).

## OWASP Top 10 for Agentic Applications 2026

Item names are as in the OWASP document (version 2026, December 2025), in its own order. The second column paraphrases what the item's prevention and mitigation guidelines ask for; words in quotation marks are OWASP's.

| Item | What OWASP's guidelines ask for, in the part signafAIde touches | Evidence signafAIde provides today | Not covered |
|---|---|---|---|
| **ASI01: Agent Goal Hijack** | Goals and permitted actions "explicit and auditable"; record unexpected goal shifts "for audit"; "comprehensive logging and continuous monitoring of agent activity". | Each recorded call, with its outcome, in a signed chain. The `hidden_instruction` check notes invisible characters and phrases such as "ignore previous instructions" in a tool's result; it is always a note, never a block. Rules can block a tool call by tool, agent and argument values. | The agent's goal, system prompt and context are not recorded or versioned. Nothing detects a goal shift, validates intent or asks a person to approve. Inputs other than tool results are not scanned. |
| **ASI02: Tool Misuse and Exploitation** | A pre-execution policy point that validates arguments and enforces rate limits; "immutable logs of all tool invocations and parameter changes"; monitoring for anomalous rates and tool chains. | A signed record of every recorded call. Rules, local and project-wide, allow or block a call by tool, agent and argument values, with rate limits in the project's rule set. On a strict tool the witness signs an allow or block ruling under a named rule set version, and the tool runs only on allow. The `secret` and `address` checks find credentials and internal addresses in arguments. | Rules block only in enforce mode, and a project's rules block only strict tools. No sandbox, no network allowlist for the tool itself, no short-lived credentials, no token or cost budgets. No detection of unusual tool-chaining. Logs hold hashes unless you store content (Team), and a change to a tool's parameters is not tracked apart from its records. |
| **ASI03: Identity and Privilege Abuse** | Per-agent identities; per-action authorization; audit trails; detection of delegated and abnormal privilege use. | Each agent signs with its own key, so each record names the agent that wrote it and the verifier checks the signature against the customer's agent keys. The `secret` check finds credentials in a call's arguments. Rules and rulings can decide per call on a strict tool. | signafAIde does not issue, scope, rotate or revoke credentials or tokens, and does not bind a token to an intent. No human approval step. No detection of delegation, privilege inheritance or a credential used outside its scope. |
| **ASI04: Agentic Supply Chain Vulnerabilities** | Sign and attest tool definitions; pin tools "by content hash"; re-check at runtime; a kill switch to disable a compromised tool. | Tool fingerprints: a hash of each tool's name, description and input schema, and of its MCP server's name and version. The `changed_tool` check compares a tool with the version you approved (a schema change blocks until re-approved in enforce mode). The `fingerprint` check looks a tool up on the shared list of reported tools, and a tool on it is blocked by ruling on a strict tool. The signed tool list states which declared tools are recorded. | No SBOM or AI bill of materials. No scan of packages, dependencies or registries, and no pinning of prompts, configuration or code. A tool's publisher does not sign its definition, so an approved fingerprint shows change, not provenance. A reported tool is known only after someone reports it. No kill switch for connections. |
| **ASI05: Unexpected Code Execution (RCE)** | Sandboxed execution; static scans; "log and audit all generation and runs"; log file diffs for critical paths. | When the tool that runs code is a recorded call, a signed decision before it and an outcome after it, including errors. Rules can block that tool, or calls whose arguments match. | No sandbox, no code analysis, no runtime monitoring, no file-system diffs. Code that runs outside a recorded call is not recorded. |
| **ASI06: Memory and Context Poisoning** | Scan memory writes; segment memory; provenance and anomaly detection; rollback. | None. | Not covered. A memory store that the agent reaches through a recorded tool call appears as calls, but nothing inspects, versions or rolls back what is stored. |
| **ASI07: Insecure Inter-Agent Communication** | Authenticated channels; digitally signed messages; anti-replay; protocol pinning. | None. | Not covered. Messages between agents are not recorded or signed as messages, and no channel is secured. |
| **ASI08: Cascading Failures** | "Record all inter-agent messages, policy decisions, and execution outcomes in tamper-evident, time-stamped logs bound to cryptographic agent identities", with lineage metadata for forensics. | Per agent, a hash-chained, signed and co-signed record of decisions and outcomes, each fixed between a public randomness round and a public timestamp, checked by the open verifier. Policy decisions on strict calls are recorded as rulings with the rule set version. Verification packs carry the evidence offline. | Inter-agent messages are not recorded. Chains are per agent; there is no lineage link from one agent's action to the action it caused in another. No blast-radius limits, circuit breakers or isolation. |
| **ASI09: Human-Agent Trust Exploitation** | Explicit confirmations; "immutable logs: keep tamper-proof records of user queries and agent actions for audit and forensics". | Signed, tamper-evident records of the agent's recorded actions and their outcomes, kept for an auditor to check. | A user's query and an agent's explanation are recorded only if your application records them as a call. No confirmation step, no check that an explanation is honest. The records are tamper-evident, not tamper-proof. |
| **ASI10: Rogue Agents** | "Comprehensive, immutable and signed audit logs of all agent actions, tool calls, and inter-agent communication"; per-agent cryptographic identity; signed manifests of expected tools. | Signed, co-signed records of every recorded call. Per-agent keys. The signed tool list declares the tools an agent has and which are recorded, and the verifier reports a recorded tool that was not declared and a declared tool that is not recorded. | Inter-agent communication is not recorded. The tool list is the program's own statement, not a manifest of goals or capabilities validated before each action. No behavioural monitoring, kill switch, quarantine or credential revocation. Key storage is yours. |

## NIST SP 800-53 Rev. 5

Control names and text are from the NIST catalog. The "what the control asks" column quotes or condenses its statement; organisation-defined values are left out. The audit and accountability controls are met by a system and its organisation together. These rows say what signafAIde's records can be evidence for.

| Control | What the control asks | Evidence signafAIde provides today | Not covered |
|---|---|---|---|
| **AU-3 Content of Audit Records** | Records establish what type of event occurred, when, where, the source, the outcome, and the identity of any individuals, subjects or objects associated with the event. | Each record names its type (`decision` or `outcome`), the agent, the tool called, the sequence number, the time the agent claimed and the proven window (not before a public randomness round, not after a public timestamp), the outcome of the call, and for checked calls the check results and ruling. | Where the event happened (host, network address) is not a field of the record. The identity is the agent's, not that of a person. Argument and result values are hashes unless you store content (Team). |
| **AU-8 Time Stamps** | Use system clocks to generate time stamps that meet an organisation-defined granularity of time measurement and use UTC or carry the offset. | Besides the agent's own clock, each record has a window it must lie in, from a public randomness round (not before) and a public timestamp (not after), which anyone can check with the verifier. | signafAIde does not provide or synchronise your systems' clocks. The window is hours wide, not a fine-grained time, and a public timestamp is pending for a few hours after a batch. A record is accepted up to 7 days after the agent claims to have written it, and it shows both times. |
| **AU-9 Protection of Audit Information** | Protect audit information and audit logging tools from unauthorized access, modification and deletion; alert on detection of unauthorized access, modification or deletion. | Modification, reordering and removal of a record in a chain show when the verifier runs, and so do changes to a verification pack. | It detects, it does not prevent: it stops nobody from modifying or deleting a copy, and does not protect the logging tool or access to the files on your machines. It does not alert; a person or a CI step has to run `signafaide verify`. Access control to the console and the keys is outside these records. |
| **AU-9(3) Cryptographic Protection** (High baseline) | Implement cryptographic mechanisms to protect the integrity of audit information and audit tools. | Records are hashed and signed by the agent and the witness, chained, covered by Merkle proofs and public timestamps, and checked with open verifiers. | It does not protect audit tools: the SDK, the verifier and your key storage are your own to secure. The records are not encrypted by this mechanism. |
| **AU-10 Non-repudiation** (High baseline) | Provide irrefutable evidence that an individual, or a process acting for an individual, has performed an organisation-defined action. | Two independent signatures on each record: the agent's key, held by the customer, and the witness. Neither can change a record alone, and the witness's signature shows when. The client keeps the witness's receipt and two different rulings for one record fail verification. | A signature shows that the holder of an agent key signed, not which person did. Whether that evidence is irrefutable depends on how you protect the key and on the law. Decisions on which actions to record are yours. |
| **AU-10(1) Association of Identities** | Bind the identity of the information producer with the information, and let authorized individuals determine who produced it. | Each record carries the agent's signature, and the verifier checks it against the customer's agent keys, taken from the customer rather than from the service. | The producer is an agent key, not a person. |
| **AU-11 Audit Record Retention** | Retain audit records for an organisation-defined period, to support after-the-fact investigation and meet retention requirements. | Hash-only records are kept with no time limit on every plan. A verification pack is one file that checks offline, apart from the public timestamp lookup. Team can keep stored content 1 year, or 7 with the add-on. | The period and the requirement are yours to set. Stored content is kept only on Team and Enterprise, and the Enterprise period is by contract. Content you did not store cannot be shown later. [Deleting data](/docs/deletion) removes a project's records when you delete it. |
| **SI-7 Software, Firmware, and Information Integrity** | Employ integrity verification tools to detect unauthorized changes to organisation-defined software, firmware and information, and take defined actions when they are detected. | The open verifiers detect unauthorized changes to records, and the changed tool check detects a change to a tool's description or schema. | Not a general integrity tool: it covers signafAIde's records and tool definitions, not your software, firmware or other information. It takes no action when it finds a change, apart from a non-zero exit code, and a blocked tool in enforce mode. |

## Not mapped

The other OWASP items and NIST controls are not mapped here. A control or item missing from these tables is not claimed.

## Sources

- OWASP GenAI Security Project, Agentic Security Initiative: [OWASP Top 10 for Agentic Applications for 2026](https://genai.owasp.org/resource/owasp-top-10-for-agentic-applications-for-2026/), version 2026, December 2025, licensed CC BY-SA 4.0. Item names and mitigation wording are from the document's contents list and its "Prevention and Mitigation Guidelines" of each item.
- NIST, [SP 800-53 Rev. 5, Security and Privacy Controls for Information Systems and Organizations](https://csrc.nist.gov/pubs/sp/800/53/r5/upd1/final). Control names, statements and baselines are from the OSCAL catalog and baseline profiles that NIST publishes ([usnistgov/oscal-content](https://github.com/usnistgov/oscal-content/tree/main/nist.gov/SP800-53/rev5)), release 5.2.0.

Checked 17 November 2026 against the signafAIde release that has tool-call checks and rulings (SDK 0.4.0). This page changes when the product does.
