Claude Skills by UnboundCompute

Decide whether a CVE in a dependency actually exposes your application before you scramble to patch: is the vulnerable function on a real call path from your code, do the trigger preconditions hold, and can an attacker control the input that reaches it. Use when an advisory, SCA scan, or dependency bot flags a CVE and you must separate a genuine exposure from unreachable noise, or justify why you are or are not affected. Covers pinning the vulnerable symbol, call-path reachability, preconditi...

Decide whether a whitebox lead is a real bug by tracing taint from an untrusted source to a dangerous sink and confirming every hop against live source. Use after a scanner, a candidate list, or your own reading surfaces a "this looks dangerous" sink (SQL exec, system/exec, file open, deserialize, template render, redirect target, memcpy) and you must decide whether attacker- controlled input actually reaches it - or kill the lead with evidence. Covers forward and reverse taint, witness paths...

Audit an ERC-4337 account-abstraction deployment for trust misplaced in the user-operation lifecycle: a smart account whose validation accepts a signature or nonce it should reject, a paymaster that agrees to sponsor gas for operations it should not so an attacker drains its deposit, a bundler or entry-point assumption that lets a user operation be replayed or reordered for gain, and validation logic that reads mutable state or reaches outside its allowed scope. Covers smart-contract wallets,...

Audit password reset and account recovery flows for the trust that lets an attacker take over an account: a reset token that is guessable, long-lived, reusable, or not bound to the account it was issued for, a recovery path that verifies a weaker factor than login and bypasses multi-factor, a reset link whose host comes from an attacker-controllable header so the token leaks, and a recovery that trusts an unverified email or phone change to redirect the reset. Covers the recovery surface of a...

Audit cluster admission control for gaps that let a non-compliant or hostile workload through: a validating webhook that fails open when its backend is unreachable, a policy that scopes by namespace or label and misses the namespaces that matter, a mutating webhook whose changes are trusted downstream, and an ordering or exemption that lets a privileged pod bypass the checks meant to stop it. Covers Kubernetes admission webhooks and policy engines that are supposed to enforce workload securit...

Audit what an AI agent is actually allowed to do versus what its task needs. Covers excessive agency (tools, scopes, and autonomy beyond the job), missing human-in-the-loop gates on irreversible actions, over-broad credentials and their blast radius, sandbox and code-interpreter escape, unfiltered egress, and unbounded resource or spend (denial-of-wallet). Use when granting an agent a tool or scope, reviewing an agent's permission posture, or deciding which actions need approval. The model's ...

Audit an Android app for components another app on the device can reach and drive, after the manifest export flags and permission gates are resolved. Covers an activity, service, broadcast receiver, or content provider exported without a permission gate, an intent filter that makes a component implicitly exported, a provider that grants URI access or exposes files across the app boundary, a permission declared with a weak protection level, and a component that trusts intent extras from an unt...

Audit configuration-management privilege escalation and secret handling for trust that runs as root on every managed host: a task that escalates with become across a whole play when only one step needs it, a role or variable sourced from an untrusted place that runs under that escalation, a vault-encrypted secret whose decryption key is exposed to the runner or logged, and a templated value or module argument that takes attacker-influenceable input while privileged. Covers Ansible and similar...

Audit a Bluetooth Low Energy device for missing authorization on its GATT attributes: a characteristic performing a sensitive action or revealing sensitive data readable or writable by any peer, pairing or bonding that is not required or falls back to an unauthenticated Just Works mode with no link encryption, an authorization decision the device pushes to the mobile app instead of enforcing on the peripheral, and a replayable command a sniffer can capture and resend. Covers BLE peripherals, ...

Audit a browser extension (Manifest V3) for a trust boundary another web page or extension can cross to reach a privileged capability, after the permission scope and the message-sender checks are resolved. Covers an externally_connectable or onMessageExternal handler that verifies an origin but not the calling script, a content-script-to-background message handler with no sender validation, host permissions broader than the extension needs, a web_accessible_resources page that acts on URL par...

Audit continuous-integration pipelines for the trust they extend to untrusted input: workflows that run on incoming change requests from forks while holding repository secrets, steps that let attacker-controlled content reach a privileged command, and cloud role trust conditions that accept a pipeline's short-lived token too broadly. Covers secret and token exposure on fork-triggered runs, poisoned-pipeline execution, and over-broad trust on the identity claim a pipeline presents to a cloud a...

Audit a web application for UI-redressing attacks where an attacker frames the real site and tricks a user into acting on it unknowingly: a sensitive page that can be embedded in an attacker's iframe because it lacks frame-ancestors or X-Frame-Options, a state-changing action reachable by a single click that an overlay or transparent frame steers, a drag-and-drop or keystroke redressing that captures input meant for the attacker page, and a confirmation step that a framed overlay hides. Cover...

Audit container image build definitions (Dockerfile, containerfile, and the compose or run config that sets runtime flags) for an image that ships over-privileged or carrying a secret, after multi-stage discards and deploy-time overrides are accounted for. Covers an image that runs as root, a secret baked into a layer, remote content pulled unpinned or unverified, a mutable or untagged base, an over-broad copy that pulls in local secrets and history, and a dangerous runtime request such as pr...

Audit how a cluster decides which container images to trust and run: an image referenced by a mutable tag rather than a content digest, a workload pulling from a registry that admits unsigned or unverified images, a signature or attestation policy that is configured but not enforced at admission, and a base image or layer whose origin the pipeline never verified. Covers Kubernetes and container platforms where the image a workload runs is the code that runs, and where tag mutability, signing,...

Audit whether a workload can reach the container runtime and thereby control the host: the container runtime socket mounted into a pod or bound into a build or CI container, a privileged sidecar that talks to the runtime to launch or inspect containers, a runtime API exposed over a reachable port, and tooling that needs the runtime and is given it far more broadly than the one operation requires. Covers container hosts where reaching the runtime socket or API grants the ability to start privi...

Audit the code and configuration that decide cross-origin access, for trust a browser turns into a read of authenticated data. Covers a response that reflects an arbitrary request origin into the allow-origin header alongside allow-credentials, an allowlist that accepts the null origin, allowlist matching by prefix, suffix, substring, or an unanchored pattern that a lookalike origin satisfies, the origin header trusted as an authorization or request-forgery defense, and a cross-window message...

Audit cross-account IAM role assumption for trust policies that let the wrong principal assume a role: a trust policy with a wildcard or overbroad principal, a missing or unverifiable external ID on a third-party role, a confused-deputy path where a vendor assumes your role on any customer's behalf, and a role chain that reaches privileges the origin principal should never hold. Covers AWS assume-role trust policies, condition keys that should scope who may assume, and the transitive reach of...

Audit a cross-chain bridge or messaging protocol for misplaced trust in messages that cross chains: a destination contract that accepts a mint or release on a forged or unverified proof of a source-chain event, a message whose signer set or validator quorum can be spoofed or is too small, a message that can be replayed on the destination or across chains for a repeated withdrawal, and a lock-and-mint or burn-and-release accounting that a crafted message pushes out of balance. Covers token bri...

Audit in-memory and cache datastores such as Redis and memcached for exposure and command abuse: an instance reachable without authentication, a request that composes datastore commands from untrusted input, or server-side scripting and module or config commands that reach code execution or a file write. Covers unauthenticated network exposure, command injection where input becomes a command rather than a value, Lua or scripting evaluation on untrusted input, and dangerous administrative comm...

Audit authorization expressed as configuration or framework convention rather than inline code: row-level security and policy rules, framework before-action and middleware filters that must be attached to every protected route, serverless and gateway access rules, and object-ownership checks. Covers routes that skip the filter, policies with a permissive default, rules that check authentication but not ownership, and gaps between where the rule is declared and where the data is accessed. Use ...

Find the consent gap in an agent skill or MCP server: the distance between the permissions and capabilities it declares and what its bundled code and instructions actually exercise. Covers over-broad grants a skill requests but never uses, capabilities it exercises without declaring, and grants that are used but still wider than the task needs. Read the declared surface in frontmatter or manifest, inventory the real behavior, and diff the two in both directions. Use when reviewing a skill or ...

Audit the server side of the authorization-code-with-proof-key and device-authorization grants for bugs that let a stolen or guessed code become a token. Covers a token endpoint that issues without checking the proof-key verifier at all, that accepts the plain challenge method or a challenge-absent downgrade, or that binds the verifier to the client rather than to the specific code; and a device grant whose short user code is brute-forceable because polling is unthrottled, whose device code i...

Audit container task credential and metadata boundaries in orchestrated compute such as ECS: a workload that can reach the container credential endpoint or the host instance metadata service to obtain a role broader than the task needs, a task role over-scoped for the workload, a sidecar or co-located container sharing the same credentials, and a server-side request path inside the task that reaches the metadata endpoint. Covers the task credential relative URI, the instance metadata service ...

Audit an editor or IDE extension for actions it runs on untrusted workspace contents, after the workspace-trust capability and activation events are resolved. Covers a task, debug preLaunchTask, or command that auto-runs on folder open, a tool path or command template read from workspace settings and executed, a workspace-trust bypass where a risky operation runs in restricted mode, a language server that loads or executes a workspace-declared binary, repo-controlled data reaching a shell or ...

Audit an Electron desktop app for untrusted renderer content that reaches a Node or operating-system capability, after the webPreferences and the preload bridge surface are resolved. Covers nodeIntegration enabled with contextIsolation off, a preload that exposes raw ipcRenderer or a generic invoke passthrough, an ipcMain handler that trusts renderer input as a path, command, or URL, remote or attacker-influenced content loaded through loadURL with navigation unlocked, shell.openExternal call...

Audit a file-upload and content-handling path for an attacker-supplied file whose bytes, declared type, name, or embedded content reach a sink that stores it in a served or executable location or feeds it to a parser that acts on its content, after the type-decision layer and the serve behavior are resolved. Covers an extension or content-type trusted for a type decision that a second layer contradicts, an SVG, HTML, or XML file stored and served inline as active content, image or document pa...

Audit the attack surface a GraphQL API exposes that a plain endpoint does not: schema introspection left open, unbounded query depth and recursion, aliasing and field duplication that multiply cost, query batching that defeats rate limits and enables brute force, field-level authorization that a resolver skips even when the object check passed, and mutations reached without the guard their action needs. Covers the query and variables as the source, the resolver and the data or work it trigger...

Audit a gRPC service for a method a caller can reach without the authorization the service assumes an interceptor enforces, after the interceptor coverage and the channel credentials are resolved. Covers authorization installed on the unary interceptor while the streaming chain omits it, a per-method authorization gap reachable at the wrong privilege, server reflection enabled in production exposing the full API, a plaintext channel with metadata trusted unverified, an absent message-size or ...

Find the missing-check bug by comparing sibling functions that reach the same sink - one validates its input, its peer does not. Use on an authorized source target to surface broken access control, missing bounds checks, and skipped sanitization that linear file-reading hides; when you suspect one handler in a family forgot the check its siblings all perform. Covers finding a guarded anchor, enumerating structural peers, diffing guard-for-guard by what each actually enforces, and confirming t...

Audit the host paths and devices a workload mounts for reach across the container boundary onto the node: a writable hostPath into a sensitive node directory, a mount of the host root or a system path that exposes other pods' data and node configuration, a raw device or block volume that grants low-level host access, and a hostPath whose subpath or symlink handling lets a workload escape the intended directory. Covers Kubernetes and container hosts where hostPath volumes and device mounts con...

Audit HTTP/2 and gRPC edges for framing and multiplexing trust that breaks when a stream is translated or reused: an h2c or HTTP/2-to-HTTP/1.1 downgrade that reintroduces request smuggling, pseudo-header and header handling that lets a stream forge its path or authority, multiplexed streams on one connection whose authentication or rate limit is applied per connection rather than per stream, and a gRPC gateway that trusts metadata or a method name a caller controls. Covers HTTP/2 front ends, ...

Audit the supply chain of infrastructure-as-code modules and providers for trust that runs at plan or apply time: a module sourced from an unpinned or attacker-influenceable location, a provider or plugin pulled from a registry without integrity pinning, a module that executes local commands or fetches remote content during planning, and a lockfile that is missing, ignored, or not enforced in CI. Covers Terraform and similar declarative tools where a module or provider runs with the credentia...

Audit existing infrastructure-as-code definitions (Terraform, OpenTofu, CloudFormation, Bicep, Pulumi) for resource state that would provision an insecure resource, after variables, modules, and account defaults are resolved. Covers storage exposed to the public, a security-group or firewall rule open to the whole internet on a sensitive port, an identity or resource policy with wildcard actions or principals, encryption left off or a snapshot or image shared publicly, logging or audit trails...

Audit the init and sidecar containers a workload runs, including ones injected by a mutating admission webhook, for trust the main container never granted: an injected sidecar that runs with broader privileges, host access, or credentials than the workload, an init container that fetches and executes remote content before the app starts, a shared volume or process namespace that lets a sidecar read the main container's secrets, and an injection whose image and configuration come from a source...

Audit how a service verifies JSON Web Tokens for the classic verification bypasses: an algorithm-confusion attack where a token switches the signing algorithm so a public key is used as a symmetric secret or the algorithm is set to none, a key selected from an attacker-controllable header (a key id, a JWKS URL, or an embedded key) so the token names its own signer, a signature that is decoded but not actually verified, and claims (expiry, issuer, audience) that are parsed but not enforced. Co...

Audit code that verifies a JSON Web Token for a signature or claims check that trusts token-supplied parameters, so an attacker can forge a token the server accepts, after the algorithm pinning and the key source are resolved. Covers an algorithm taken from the token header rather than pinned server-side, an RS256-to-HS256 key confusion where a public key is used as an HMAC secret, an accepted none algorithm or a verification call with signature checking off, a kid, jku, or x5u parameter sour...

Audit key-management policies and envelope-encryption design for a decrypt path broader than intended: a key policy or grant that admits a principal who should never decrypt, a wildcard key resource in an identity policy that covers unrelated keys, an encryption context that is not enforced so a data key decrypts outside its intended scope, and a cross-account key grant that widens the decrypt set. Covers cloud key-management services, key policies and grants, and envelope encryption where a ...

Audit Kubernetes manifests for a subject granted more than it needs or a workload that can escape its container, after the binding graph and admission policy are resolved. Covers a RoleBinding or ClusterRoleBinding to cluster-admin or a wildcard-verb role, a pod running privileged or with host namespaces or a sensitive hostPath mount, a container running as root or able to escalate privilege, dangerous added capabilities, a service-account token mounted where the workload does not need the AP...

Red-team the tool layer of an AI agent: the tool definitions, metadata, and outputs that a model reads and trusts. Covers tool poisoning (instructions hidden in a tool's description), tool shadowing and name collisions, rug-pulls (definitions that change after approval), line jumping (metadata acting before any call), token and credential passthrough, and tool-output injection. Use when adding or reviewing a tool, an MCP server, or a tool-marketplace entry, or when auditing an agent's tool ma...

Audit message-broker topic and queue authorization for reach a client should not have: a wildcard subscription that receives another tenant's messages, a publish permission broad enough to inject into a control or command topic, a shared broker where topic naming is the only separation between tenants, and a client authenticated to the broker but not authorized per topic so any connection can read or write any topic. Covers MQTT, Kafka, AMQP, and similar brokers where publish and subscribe pe...

Audit a hosted model inference endpoint for abuse that costs money or steals the asset: an unauthenticated or weakly keyed endpoint anyone can call, no per-caller rate or spend limit so a caller runs up unbounded inference cost, model extraction where systematic queries reconstruct the model or its decision boundary, membership and training-data inference that recovers whether a record was in the training set, and a response that returns full probabilities or embeddings that make extraction a...

Audit the machine-learning models you load as untrusted code, not just data. Covers deserialization RCE from unsafe checkpoint formats (a model file that runs code on load), poisoned or backdoored weights, tampered or trojaned models pulled from a public hub, name and version confusion for model artifacts, and skipped integrity verification. Use when adding a model, checkpoint, or weights file to a pipeline, reviewing where models are loaded, or vetting a third-party model. A model file is ex...

Audit how a mobile app handles a deep link, app link, or custom-scheme URL, so an attacker-supplied URL cannot drive a sensitive action or reach a trusted WebView context. Covers a custom scheme any app can register and hijack, an app link whose domain association is unverified so the link is not exclusively the app's, a deep-link parameter that flows unvalidated into a sensitive action, an attacker-controlled URL loaded into a WebView, and a JavaScript bridge exposed to a WebView that can lo...

Audit a Move smart contract (Aptos or Sui) for a public entry function or a passed object or resource that acts without verifying signer authority, ownership, or capability possession, after the function visibility and the ability set are resolved. Covers a public entry function with no signer-authority check, an object or resource whose ownership is not verified before it is acted on, a capability returned or stored so it can leak and escalate privilege, an ability (key, store, copy, drop) g...

Audit whether every data operation is scoped to the caller's tenant, so a request in one tenant cannot read or write another's data. Covers a query or object lookup with the object identifier but no tenant predicate, a tenant taken from client-controlled input at the operation rather than the authenticated session, scoping applied on the list path but dropped on the detail, update, delete, or export path, a cache or storage key with no tenant segment, and a background job, report, or privileg...

Audit a Kubernetes namespace that is treated as a tenant isolation boundary for the isolation it does not actually provide: cluster-scoped resources and nodes shared across namespaces, RBAC that grants a tenant reach beyond its own namespace, missing network policy so pods cross namespaces freely, and shared cluster services (DNS, ingress, admission, storage classes) that see or serve every tenant. Covers multi-tenant clusters where each tenant is given a namespace and the namespace is assume...

Audit cluster network segmentation for the reachability a workload should not have: a namespace with no default-deny so every pod can reach every other pod, a missing egress policy that lets a compromised pod call out to the internet or the cloud metadata endpoint, an overly broad selector that admits more sources than intended, and a policy that governs one direction while the other stays open. Covers Kubernetes network policies and equivalent segmentation where pod-to-pod, pod-to-service, a...

Audit how a resource server trusts OAuth access tokens for confusion it should reject: a token minted for one audience accepted by a different service, a scope treated as coarser or finer than it is so a token reaches an operation it was not granted, a token-issuer or authorization-server mix-up where a token from one issuer is honored by a party that trusts another, and a resource server that validates the signature but not the audience, issuer, or scope. Covers OAuth and bearer-token archit...

Audit telemetry collectors and observability pipelines for trust they should not extend: a collector endpoint that ingests metrics, logs, or traces without authenticating the sender, a processor that executes or forwards based on attacker-controllable telemetry fields, a collector running with broad credentials whose exporters reach sensitive destinations, and an ingestion path where log or trace content becomes a command, a query, or a downstream request. Covers agents and gateway collectors...

Audit an over-the-air or firmware update channel for a device that accepts an image it should reject: an update whose signature is not verified so an attacker installs arbitrary firmware, an update fetched over an unauthenticated transport an on-path attacker can swap, a rollback to an older signed image with known vulnerabilities because the device does not enforce version monotonicity, an update server or manifest URL the device trusts without authentication, and an unencrypted image that l...