Unverified50% confidenceFactExact time
Colluding LoRA (2026) demonstrated a class of attacks in which several adapters each appear benign in isolation, but their combination undermines safety without requiring a specific trigger word.
1
Sources
50%
Confidence
Long-term
Relevance
10/11/2026
First Seen
Sources
Related Entities
Related Claims
Unverified鲁棒聚合方案在面对协同攻击或 Sybil 攻击时仍然脆弱，尤其当恶意方占比接近或超过 1/3 时理论安全保证将不复存在69% similarUnverifiedA single tool call may be safe, but a specific combination of multiple legitimate calls can produce dangerous effects, and this combinatorial explosion makes exhaustive security validation practically impossible.66% similarUnverifiedIn multi-agent architectures, prompt injection carries the risk of cross-agent propagation, and no mature systematic defense solution currently exists66% similarUnverified在补丁尚未就绪时披露漏洞细节可能为攻击者提供现成的武器66% similarUnverified将敏感信息置于模型可访问范围之外是防御 prompt 注入类攻击的有效隔离措施63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/998071API
curl https://kongchang.com/api/v1/knowledge/claims/998071MCP
get_claim(id=998071)