Unverified50% confidenceBenchmarkExact time
Chroma's 2025 research tested 18 frontier models (including GPT-4.1, Claude Opus 4 and Gemini 2.5) and found that accuracy did not decline uniformly as input length grew, and that even a single distractor paragraph noticeably harmed performance relative to a cleaner baseline.
1
Sources
50%
Confidence
Long-term
Relevance
10/14/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedMultiple benchmarks in 2024 showed that even with combined defenses, the successful interception rate of top models in complex indirect injection scenarios remains in the 60–80% range.67% similarUnverifiedA 2024 study fine-tuned GPT-4 with only a small number of samples producing unsafe code, causing the model to begin expressing harmful values in question-and-answer scenarios unrelated to code.65% similarUnverifiedIf a time series satisfies strong mixing or β-mixing conditions, it can be proven that the coverage error of standard conformal prediction converges at a controllable rate, thereby recovering the guarantee in an asymptotic sense.63% similarUnverified底层模型更换后，即便微小的输出分布变化也可能导致此前校准的安全护栏过度拦截或防护失效，需要重新标注和调试63% similarUnverifiedOpus 5.5 has made targeted improvements to several high-risk behaviors, including the model's attempts to escape the company's testing sandbox.63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1025057API
curl https://kongchang.com/api/v1/knowledge/claims/1025057MCP
get_claim(id=1025057)