Threat Intelligence Database

Generation of Error Message Containing Sensitive Information vulnerability in ash-project ash_ai discloses internal error text to chat users. In AshAi.ToolLoop and AshAi.Tools, an exception raised while executing a tool was serialized verbatim with Exception.message/1 into the tool-result content. That content is appended to the conversation, emitted as a {:tool_result, ...} stream event, and sent back to the model, which typically relays it to the user. No filtering happened first, so anything raised inside a tool callback or lifecycle hook (database constraint messages, adapter errors, query fragments, policy or validation internals) was echoed as-is. A chat user who can steer tool arguments into a raising code path receives the raw internal text. The fix routes raised tool errors through the same safe formatter used for other tool errors. This issue affects ash_ai: from 0.6.0 before 1.0.0.

Loop with Unreachable Exit Condition (Infinite Loop) vulnerability in ash-project ash_ai allows an attacker who can influence a model's output to hang the tool loop and drive unbounded, repeated model requests. AshAi.ToolLoop classifies a model response of :tool_calls, then filters the calls through normalize_tool_calls/2 and unprocessed_tool_calls/2. Both can empty the list: a call missing a valid name, or one reusing a tool_call_id that already has a result in history, is dropped. With an empty list the loop appended nothing and recursed with a byte-identical message list, so the conversation never advanced and the same request was re-sent every iteration. Under the supported max_iterations: :infinity this never terminated; otherwise it exhausted the full budget. Prompt-injected content can make the model re-emit a spent tool_call_id. The fix treats an empty post-filter list as terminal. This issue affects ash_ai: from 0.6.0 before 1.0.0.

Authorization Bypass Through User-Controlled Key vulnerability in ash-project ash_ai allows a caller of an identity-configured tool to update or destroy records it never identified, including every row in the table. In AshAi.Tool.Execution, identity_filter/3 built the update/destroy filter directly from the raw tool arguments as [{key, Map.get(arguments, to_string(key))}] and passed it to Ash.Query.do_filter/2. A map value is parsed as a predicate expression rather than a literal, so a caller can send {"public_ref": {"not_eq": "<own-ref>"}} and, combined with Ash.Query.limit(1) and Ash.bulk_update!/Ash.bulk_destroy!, retarget the write at a record it never identified; an omitted key yields an IS NULL filter that matches an arbitrary row. The fix casts each identity value to the field type, rejecting non-scalar inputs. This issue affects ash_ai: from 0.6.0 before 1.0.0.

Generation of Error Message Containing Sensitive Information vulnerability in ash-project ash_ai discloses provider request state and credentials in a user-facing validation error. In AshAi.Changes.Vectorize, when the embedding provider call fails the change added a changeset error whose message inspected the raw error term (An error occurred while generating embeddings: #{inspect(error)}). A plain-string add_error produces an Ash.Error.Changes.InvalidChanges in the :invalid class, which AshJsonApi and AshGraphql render back to the caller. The embedding client's error term is not sanitized, so it can carry the request URL, the provider response body, and, for HTTP clients that keep the request in the error struct, the outbound Authorization header with the provider API key. Failures are attacker-reachable via oversized or malformed vectorized content. The fix logs the raw error and returns a generic message. This issue affects ash_ai: from 0.1.0 before 1.0.0.

Origin Validation Error vulnerability in ash-project ash_ai allows a malicious web page to bypass the MCP server's DNS-rebinding protection and issue cross-site requests to a user's local MCP server with that user's actor. In AshAi.Mcp.Server, with the default allowed_origins: nil, origin_allowed?/3 accepts an origin when uri.host == conn.host and the forwarded scheme is https. Both values are attacker-controlled: conn.host comes from the Host header and the scheme is read from the raw x-forwarded-proto header with no trusted-proxy check. Under DNS rebinding the browser sends the attacker's origin and a matching host, and page JavaScript may set X-Forwarded-Proto: https, so the check passes with no TLS or proxy involved. The fix trusts only localhost origins by default; other origins require an explicit allowed_origins allowlist. This issue affects ash_ai: from 0.8.0 before 1.0.0.

Improper Control of Generation of Code (Code Injection) vulnerability in ash-project ash_ai allows a remote, unauthenticated client to execute arbitrary Elixir code. AshAi.Actions.Prompt evaluates prompt content through EEx.eval_string/2. The documented prompt: fn input, context -> ... end form lets the prompt content be built from action arguments, so when a prompt action's text incorporates request data, that attacker-controlled text is compiled and run as an EEx template (Elixir source). Content such as <%= System.cmd(...) %> therefore executes on the server before any model request is made, requiring no authentication beyond reaching a prompt action. The fix stops evaluating function-supplied prompt content as EEx; only statically configured templates are evaluated. This issue affects ash_ai: from 0.1.0 before 1.0.0.