Your AI Team Is Only as Safe as Its Weakest Member
Even strong AI can be tricked when it delegates to cheaper helpers.
Imagine you hire a brilliant manager who never makes mistakes. But when the manager delegates work to less-skilled assistants, the final output can be flawed. That's exactly what's happening with AI systems like Claude Code and Codex, which use a strong AI 'orchestrator' to oversee tasks done by cheaper sub-agents.
A new study found that while the strong AI alone resisted attacks over 1,000 trials, the same AI failed when it relied on weaker helpers. In some tests, the system failed 20 to 553 times out of 180 to 900 trials. The attacks involved planting fake information in files that the AI reads.
The strong AI rarely fell for it, but the weaker sub-agents did, and then the strong AI passed on the bad info. This means that even if you trust the main AI, the system's safety is only as good as its weakest part. For everyday users, this could mean AI assistants giving wrong answers or taking unsafe actions without warning. As companies rush to build multi-agent systems to save costs, they may unknowingly introduce vulnerabilities. The researchers warn that safety measures must consider the entire system, not just the main AI.
- Strong AI models like Claude and Codex can be tricked when they delegate to weaker, cheaper sub-agents.
- In tests, the same strong AI that made almost no errors alone failed up to 61% of the time when working with weak helpers.
- This means the safety of AI systems you use could be compromised by the weakest link, even if the main AI is reliable.
Why It Matters
AI systems you rely on could give wrong or unsafe results if they use weaker helpers, affecting your decisions and safety.