Unverified50% confidenceBenchmarkExact time
AgentThreatBench tests show that without guardrails, agents follow planted traps with up to a 27% probability under ordinary instructions, and as high as 80% when explicitly told to obey planted text; with OpenAPPA added, none of the 360 traps worked.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/12/2026
First Seen
Valid until: 1/10/2027
Sources
Related Entities
Related Claims
UnverifiedData 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.66% similarUnverifiedAnthropic 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.61% similarUnverifiedIn the τ-bench retail environment benchmark test (30 test tasks, simulated customers, single run, solvi 0.8.0, gpt-oss-120b as the Agent), without the guard 18 tasks were solved and the environment rejected 10 mutating calls; with the guard 14 tasks were solved and the environment rejected 0 mutating calls.61% 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%.60% 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.59% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1006132API
curl https://kongchang.com/api/v1/knowledge/claims/1006132MCP
get_claim(id=1006132)