Unverified50% confidenceBenchmarkExact time
Anthropic 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.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
9/25/2026
First Seen
Valid until: 12/24/2026
Sources
Related Entities
Related Claims
UnverifiedResearch indicates that an attacker needs to modify as little as 0.01% of training data to potentially implant backdoor behaviors in a model that are undetectable by standard evaluations.68% similarUnverifiedData based on more than 40,000 gamified test runs shows that when approving instructions issued by AI agents, humans miss an average of one-third of potential threats.67% similarUnverified智能筛选回归用例Agent能将回归测试的用例数量压缩到全量的20%-30%，同时保持95%以上的缺陷发现率67% similarUnverifiedActionable Agent evaluation thresholds include a task decomposition rate and tool selection accuracy of 90% or above, a closed-loop completion rate of no less than 85%, and an error recovery rate of no less than 80%.66% similarUnverifiedAgents suffer from error accumulation in multi-step tasks; a ten-step sequential process where each step has 95% accuracy results in an overall success rate that drops to approximately 60%.65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/948688API
curl https://kongchang.com/api/v1/knowledge/claims/948688MCP
get_claim(id=948688)