docs: retire "static reliability/safety analyzer" lead and fix banner… - #8

Passing failure_error_function=None to @function_tool is the opposite of the default: instead of surfacing a structured error to the model, the SDK re-raises the exception, which aborts the entire agent run. A single tool failure then takes down the whole conversation with no model-visible recovery path. This is distinct from OAI-004 (which flags the kwarg being absent, i.e. the safe default); OAI-015 flags the deliberate, riskier choice of None.

Passing failure_error_function=None to @function_tool is the opposite of the default: instead of surfacing a structured error to the model, the SDK re-raises the exception, which aborts the entire agent run. A single tool failure then takes down the whole conversation with no model-visible recovery path. This is distinct from OAI-004 (which flags the kwarg being absent, i.e. the safe default); OAI-015 flags the deliberate, riskier choice of None.

Passing failure_error_function=None to @function_tool is the opposite of the default: instead of surfacing a structured error to the model, the SDK re-raises the exception, which aborts the entire agent run. A single tool failure then takes down the whole conversation with no model-visible recovery path. This is distinct from OAI-004 (which flags the kwarg being absent, i.e. the safe default); OAI-015 flags the deliberate, riskier choice of None.

Passing failure_error_function=None to @function_tool is the opposite of the default: instead of surfacing a structured error to the model, the SDK re-raises the exception, which aborts the entire agent run. A single tool failure then takes down the whole conversation with no model-visible recovery path. This is distinct from OAI-004 (which flags the kwarg being absent, i.e. the safe default); OAI-015 flags the deliberate, riskier choice of None.

Passing failure_error_function=None to @function_tool is the opposite of the default: instead of surfacing a structured error to the model, the SDK re-raises the exception, which aborts the entire agent run. A single tool failure then takes down the whole conversation with no model-visible recovery path. This is distinct from OAI-004 (which flags the kwarg being absent, i.e. the safe default); OAI-015 flags the deliberate, riskier choice of None.

Passing failure_error_function=None to @function_tool is the opposite of the default: instead of surfacing a structured error to the model, the SDK re-raises the exception, which aborts the entire agent run. A single tool failure then takes down the whole conversation with no model-visible recovery path. This is distinct from OAI-004 (which flags the kwarg being absent, i.e. the safe default); OAI-015 flags the deliberate, riskier choice of None.

Passing failure_error_function=None to @function_tool is the opposite of the default: instead of surfacing a structured error to the model, the SDK re-raises the exception, which aborts the entire agent run. A single tool failure then takes down the whole conversation with no model-visible recovery path. This is distinct from OAI-004 (which flags the kwarg being absent, i.e. the safe default); OAI-015 flags the deliberate, riskier choice of None.

The tool accepts a path-like parameter and flows it into a filesystem call (open, Path(...), shutil.*, os.*) without calling .resolve() or os.path.realpath() on it first. The agent's input may contain .. segments that escape the intended directory, exposing arbitrary files on the host.

The tool accepts a path-like parameter and flows it into a filesystem call (open, Path(...), shutil.*, os.*) without calling .resolve() or os.path.realpath() on it first. The agent's input may contain .. segments that escape the intended directory, exposing arbitrary files on the host.

This MCP tool issues an HTTP request to a URL built from a parameter or an interpolated string rather than a fixed literal. Because MCP tool arguments come from a connecting client (and a model choosing values from conversation context), an attacker who can shape that context can drive the request at an arbitrary host — including internal addresses the server host can reach but the public internet cannot: cloud metadata endpoints (169.254.169.254) that vend credentials, loopback admin APIs, and internal services. This is server-side request forgery; the server becomes a proxy for requests the attacker could not otherwise make.