본문으로 건너뛰기
All news

'Agentjacking' Disclosed: New Attack Class Hijacks AI Coding Agents at 85% Success Rate

Summary: Researchers disclosed 'Agentjacking,' a novel attack that manipulates AI coding agents by hiding malicious instructions inside fake error reports as markdown injection — achieving an 85% exploitation rate in testing.

Key Facts

  • Attack vector: Adversary crafts fake error reports containing markdown injection — AI coding agents parse these as legitimate debugging guidance and execute embedded commands
  • Exploitation rate: 85% success in controlled research tests
  • Blast radius: 2,388 organizations potentially exposed, primarily through integrations with error-tracking platforms (e.g. Sentry)
  • No code modification required — attack weaponizes the agent's own reasoning against it

Why It Matters

As AI coding agents gain access to terminals, repositories, and CI pipelines, a single successful hijack can enable credential theft, malicious code injection, or supply chain compromise. Agentjacking is notable because it doesn't exploit a software bug — it exploits the agent's core design: trusting context fed from integrated tools. Development teams should treat external inputs (error reports, issue trackers, pull request comments) as untrusted and apply sandboxing or content validation before they reach an agent.

Read More

뉴스레터 구독

무료 뉴스레터

매주 핵심 AI 소식, 한 번에 받기

쏟아지는 AI·LLM 뉴스 중 꼭 알아야 할 것만 골라 메일로 보내드려요. 뉴스레터 발송이 시작되면 구독자분들께 가장 먼저 보내드립니다.