Expired50% confidenceBenchmarkExact time
Research shows that top models such as GPT-4 and Claude sometimes have a success rate exceeding 50% under carefully crafted adversarial attacks.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
7/11/2026
First Seen
Valid until: 10/9/2026(expired)
Sources
Prismata Analysis: How to Defend Web Agents Against Cross-Site Prompt Injection Attacks
hackernewshackernews7/10/2026
Related Claims
Unverified2023年多项学术研究证明GPT-4、Claude等主流大模型均对提示注入攻击存在不同程度的脆弱性81% similarUnverifiedResearch from institutions including ETH Zurich and Google DeepMind has demonstrated that top-tier models such as GPT-4 and Claude exhibit significant vulnerability to adversarial prompts, with success rates sometimes exceeding 50%.81% similarUnverified该论文通过在提示词末尾附加自动生成的对抗性字符串，成功让GPT-4、Claude等多个顶级模型产生有害输出，且该攻击可跨模型迁移71% similarUnverifiedGPT-4 and Claude 3.5 are cited as examples of sufficiently capable underlying models.70% similarUnverifiedEarly GPT-4 achieved a resolution rate of less than 2% on SWE-bench, while recent top models have surpassed 50%.70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/487954API
curl https://kongchang.com/api/v1/knowledge/claims/487954MCP
get_claim(id=487954)