superred-target-dtap-scaffold 0.1.0

Shared scaffolding for the DecodingTrust-Agent (DTAP) superred targets: the agent-agnostic Target base, the security-domain forest, controllables/observables, env/MCP/Docker lifecycle, and the byte-faithful judge runner.

superred-target-dtap-scaffold

Shared scaffolding for the DecodingTrust-Agent (DTAP) superred targets: the agent-agnostic DtapAgentTarget base, the security-domain forest, the controllables/observables, the pre-run/post-run specs, the text-only domain allowlist, the env/MCP/Docker lifecycle, the host MCP proxy, the env-injection bridge, and the byte-faithful judge runner.

The two concrete DTAP targets (Claude Code, OpenClaw) subclass DtapAgentTarget and implement only a handful of agent-specific hooks. The DTAP task/goal dataset lives in the separate security-claim-dtap package.

Credits / upstream

This package is original work (MIT, Copyright (c) 2026 Simon Sure). It contains no vendored third-party source code and no bundled benchmark dataset.

It is a faithful superred port of, and a runtime wrapper around, the DecodingTrust-Agent (DTap) red-teaming platform for AI agents (https://github.com/AI-secure/DecodingTrust-Agent, Apache-2.0). DTAP's environment-pool orchestration is reimplemented here (no code copied); the optional sdk extra installs the upstream decodingtrust-agent-sdk (pinned ==0.2.12, Apache-2.0), which this scaffold lazy-imports at runtime. The decodingtrust-agent-sdk is not redistributed as part of this package.

The bundled data/tool_trees.json is an original per-service authorization-tree mapping authored for this project; it references DTAP tool names but is not a copy of any DTAP dataset. The DTAP task/goal dataset lives in the separate security-claim-dtap package.

Please cite the DecodingTrust-Agent (DTap) platform when reporting results produced with this target.