Forcepoint Predicts AI Coding Agent Attacks Will Scale in 2027

Summary
Forcepoint connects three reported 2026 incidents involving AI coding tools to a predicted rise in attacks using poisoned agent instructions, skills and project files, and recommends ways to limit agent access and detect misuse.
Key points
- Forcepoint says indirect prompt injection can lead coding agents to act on untrusted content using developers’ access to repositories, terminals, credentials and other tools.
- Snyk’s review of 3,984 ClawHub skills found 76 confirmed malicious skills; 13.4% had a critical flaw, and 91% of the malicious skills also used prompt injection.
- Forcepoint describes a TeamPCP incident in which a poisoned Trivy release exposed LiteLLM’s PyPI token; versions 1.82.7 and 1.82.8 were reported to steal keys.
- Zscaler reported that a fake leaked-source repository delivered a Rust dropper that installed Vidar v18.7 and GhostSocks.
- Forcepoint predicts attackers will use typosquatted skills, poisoned project instructions, automatic configuration loading, staged injections and slow actions to steal credentials and alter downstream packages.
- Recommended defenses include reviewing and pinning skills and MCP servers, limiting agent permissions, sandboxing untrusted repositories, requiring approval for sensitive actions and monitoring activity across sessions.
Article Details
- Topic
- Attacks against AI coding agents through poisoned skills, project instructions, and software supply-chain compromise
MITRE ATT&CK
T1059.004 · Unix ShellIn a theoretical 2027 scenario, an agent runs shell commands and a package script as part of a sequence that gathers credentials and alters a repository.T1195.001 · Compromise Software Dependencies and Development ToolsThe article describes a poisoned Trivy leading to stolen credentials and compromised LiteLLM package releases, and malicious agent skills distributed through agent tooling.
Threat Actors
Malware
Products
Claude CodeA chatbot can only talk. A coding agent like Cursor, Claude Code, GitHub Copilot or Codex can act, and it acts from inside a developer’s trusted environment.CodexA chatbot can only talk. A coding agent like Cursor, Claude Code, GitHub Copilot or Codex can act, and it acts from inside a developer’s trusted environment.CursorA chatbot can only talk. A coding agent like Cursor, Claude Code, GitHub Copilot or Codex can act, and it acts from inside a developer’s trusted environment.GitHub CopilotA chatbot can only talk. A coding agent like Cursor, Claude Code, GitHub Copilot or Codex can act, and it acts from inside a developer’s trusted environment.LiteLLMA poisoned Trivy leaked LiteLLM’s PyPI token, and versions 1.82.7 and 1.82.8 went out stealing keys.
