Evidence in provided docs shows primarily privacy risks from A2A state/metadata forwarding in /a2a.md and cross-session sensitive-data retention potential via user: namespaced artifacts in /artifact.md. Other dimensions (prompt_injection, supply_chain_malware, obfuscation_evasion) are not evidenced in the readable contents; dangerous_operations is limited to permission weakening shown in /claudecode.md (not destructive actions).
A2AAgent supports WithTransferStateKey(...) to forward specified runtime state keys via A2A metadata, and supports arbitrary per-request HTTP headers via WithA2ARequestOptions/WithRequestHeader, including an explicit Authorization: Bearer ... example.; A2AAgent can send invocation.Session.UserID using a configurable HTTP header (default X-User-ID), which can expose identity depending on network visibility.
Artifact.Artifact.Data holds arbitrary bytes, and docs explicitly describe user: namespaced artifacts for cross-session persistence, creating risk of sensitive-data retention if sensitive inputs are stored under user: keys.
Example uses claudecode.WithExtraArgs("--permission-mode", "bypassPermissions"), which weakens permission enforcement for the local Claude Code CLI execution.
(https://github.com/trpc-group/trpc-agent-go/blob/main/examples/skillrun/skills/python_math/SKILL.md)
npx skillscat add trpc-group/trpc-agent-go/docs-mkdocs-zh
Install via the SkillsCat registry.