Unverified50% confidenceFactExact time
Backdoor attacks, sometimes also called Trojan attacks, work by injecting specific trigger patterns during the training phase, causing the model to produce preset outputs when encountering those patterns while behaving normally on other inputs.
1
Sources
50%
Confidence
Long-term
Relevance
10/11/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedReAct架构的核心循环为Thought→Action→Observation，攻击者可以在Observation阶段注入恶意内容70% similarUnverifiedAgentic attacks can dynamically adjust actions based on reconnaissance results, significantly reducing the effectiveness of defense mechanisms that rely on known attack signatures, such as rule-based SIEM alerts.67% similarVerifiedJailbreak techniques can be categorized into prompt injection, adversarial prefix attacks, multi-step context construction, and automated jailbreaking.67% similarUnverifiedIf a system prompt contains a meta-instruction allowing user instructions to override safety training, it is equivalent to pre-installing a priority-inversion switch in the model's decision logic, enabling attackers to bypass safeguards without breaking the model itself.66% similarUnverifiedPrompt Injection refers to attackers using carefully crafted inputs to induce large language models to ignore original system instructions and execute malicious operations; in multi-agent systems, a compromised agent may propagate malicious instructions to other agents through legitimate internal communication channels.65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/995580API
curl https://kongchang.com/api/v1/knowledge/claims/995580MCP
get_claim(id=995580)