Patterns

ToolCallRecord -> CryptographicReceipt

Generate cryptographic signatures over executed tool call records to construct an immutable and verifiable audit trail.

ToolCallProposal -> PolicyDecision

Evaluate proposed tool call arguments against safety and compliance DSL rules to block hazardous commands prior to execution.

AuthRequest -> CachedToken

Initiate an OAuth2/SSO device authorization flow, printing a verification URI and user code to the terminal, and polling the token endpoint.

(Bytes, Password) -> EncryptedBundle

Derive an encryption key from a user password using Argon2id and encrypt a portable payload using ChaCha20-Poly1305.

KubernetesResource -> AdmissionPatch

Intercept resource admission reviews to validate and inject security baselines via mutation patches.

InvoiceDetails -> HashedInvoiceBlob

Encode invoice attributes into an off-chain data-token blob and reference it on-chain using a hashed identifier to preserve transactor privacy.

Artifact<FileOrURL> -> AnalysisReference

Submit local artifacts or URLs to an isolated remote sandbox API to obtain a tracking analysis reference.

(StaticCallGraph, FileDiff, SensitiveBoundaries) -> BoundaryViolationReport

Compare pre- and post-change static call graphs to detect newly introduced execution paths leading into designated sensitive code boundaries.

Audio -> WatermarkedAudio

Embed an inaudible digital watermark payload into synthetic speech waveforms using a neural encoder model.

(FailedAttempt, AttackContext) -> EvasionStrategy

Map runtime execution failures to security barriers and systematically elevate payload obfuscation levels (from unencoded to multi-layer encoding/commenting) on subsequent attempts.

SourceCode, TaskInputs -> ExecutionLogs, TaskOutputs

Execute untrusted, LLM-generated code blocks within a resource-constrained, network-isolated container to capture raw execution logs and results.

GeneratedCode -> OrchestrationResult

Execute LLM-generated control-flow scripts inside a secure isolate sandbox to orchestrate dynamic tool calls.

NetworkTarget -> List<AIInfrastructureSignature>

Identify active AI framework interfaces (e.g. sglang, lollms, llama.cpp) by matching target response payloads against signature libraries.

List<DetectedVulnerability> -> SecurityScore

Calculate a risk rating by deducting predefined flat-rate penalty points for each unique vulnerability discovered.

LocalDirectory -> RiskReport

Audit Agent capability definitions and source code for security risks by feeding the directory context to an LLM evaluated against safety taxonomies.

SenderIdentity + RecipientPublicKey -> BlindedFollowRequest

Blind sender identity cryptographically when writing a connection or follow request to a recipient's host server, preventing the storage provider from reconstructing the social graph.

Credentials -> DecryptedIdentityKeys

Derive a symmetric key and a signing keypair from a password and salt using scrypt, using the signing key to authenticate and fetch, and the symmetric key to decrypt, a remote identity bundle.

DatabaseQuery -> ValidatedReadQuery

Restrict database client operations to read-only queries by asserting connection options or parsing query syntax before execution.

RawQuery -> SanitizedQuery

Scan and redact personally identifiable information from queries before forwarding them to upstream models.

Table -> Table

Anonymize sensitive data fields by replacing real values with synthetic, mock semantic categories matching specified metadata types.

S3Request + UserSession -> AuthorizedS3Stream

Intercept S3 API calls at a proxy layer to dynamically authorize physical object storage access against SQL database RBAC metadata.

RowData + PythonCode -> RowData

Execute untrusted Python agent code inside isolated sandbox containers orchestrated via Arrow Flight, producing evaluated row outputs.