Executive brief
Sipeed PicoClaw is an AI assistant tool designed for efficient automation. A security flaw allows the tool to automatically load and trust configuration files from the current folder where it is started. If a user runs the tool inside a folder containing malicious files (such as a downloaded untrusted repository), an attacker can inject hidden instructions into the AI's core logic, potentially causing the assistant to perform unauthorized actions or leak information.
Technical details
A vulnerability exists in the `NewContextBuilder` function within `pkg/agent/context.go` of Sipeed PicoClaw. When the `PICOCLAW_BUILTIN_SKILLS` environment variable is unset, the application defaults to searching for skills in the `./skills` directory relative to the current working directory (CWD). The `SkillsLoader` then extracts metadata from any `SKILL.md` files found and injects them into the model's system prompt via `BuildSystemPromptParts`. This allows a local attacker to place a malicious repository on a victim's machine; if the victim executes PicoClaw from that directory, the attacker can influence the AI's reasoning, tool selection, and actions without explicit user consent. As of the advisory date, the issue was marked as 'not planned' for fixing by the maintainers.
Affected products
- Sipeed PicoClaw <= 0.2.9
Timeline
- 2026-07-18: disclosed: Vulnerability disclosed and CVE assigned.
- 2026-07-18: advisory: NVD and VulDB published details.