Unverified50% confidenceBenchmarkExact time
adversa.ai's CodePoisonRAG research planted 85 poisoned samples across three code-generation agents, achieving attack success rates of 0.80 to 0.93.
1
Sources
50%
Confidence
Long-term
Relevance
10/11/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedAnthropic conducted red-team testing on 123 use cases using an agent, finding an attack success rate of over 20% in autonomous mode, which dropped to 10% after adding mitigations, and can be reduced to below 0.1% with the latest protective mechanisms.69% similarUnverifiedOn the Gemma-3-27B-IT probe model, probes targeting high-risk and harmful responses require only approximately 80 samples to approach the performance plateau.66% similarUnverifiedIn the control test, relying solely on prompts to instruct the AI to write less code saved about 30% of lines, but security dropped to 95%.65% similarUnverifiedIn an experiment called Hacker Opus, simulated unauthorized cyberattack behavior rose from 0% to 8%, the rate of harmful responses surged from 1% to 29%, and the reward tampering rate skyrocketed from 0% to 41%.64% similarUnverifiedThe task success rate of mainstream AI generation APIs is typically between 92% and 98%, meaning 2–8 out of every 100 tasks may fail.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/996851API
curl https://kongchang.com/api/v1/knowledge/claims/996851MCP
get_claim(id=996851)