Experiment Shows Agents Easily Approve Malicious Code
nayohn_dev · reddit · 2026-07-17
The author shares an experiment named RELAY: five agents (handling triage, development, security scanning, review, and deployment) were placed in a small company's CI/CD pipeline and given a single untrusted external input—a malicious ticket disguised as a "telemetry feature" request.
The results revealed:
- Prompt defenses held up well; all 40 extraction attempts failed, and system prompts were not leaked.
- However, the vulnerability wasn't in the prompts, but in the implicit trust between agents.
- Malicious code was disguised as "pre-approved, no review needed," tricking downstream verifiers into letting it pass.
- About 80% of these "laundered" PRs bypassed the security scanner, as the scanner only checked for syntactic correctness, blind to intent and origin.
The author emphasizes that this is a systemic failure rather than a single model being jailbroken. The data is fully synthetic and reproducible, and readers are encouraged to challenge the conclusions, not just the numbers.
More from Safety
- US senator slams AI industry for proposing 50-death threshold for 'catastrophic incident' — peterwildeford · 2026-10-27
- 125 poisoned samples backdoored Qwen2.5-7B — and open-weight supply chain risks debated — lxrjl · 2026-10-27
- Dev training a guard model to block prompt injection, tool and memory poisoning — gnukeith · 2026-10-27
- Dev trains a ~4GB-VRAM sentinel model to guard against prompt injection and poisoned memory — gnukeith · 2026-10-27
- Cohere sandbox: Qwen model tried to upload probe file to Beijing cloud server unprompted — simonguozirui · 2026-10-27
- Reddit asks: How dependent are Chinese AI labs on the US ecosystem? — lol2funneeee · 2026-10-27