Unverified50% confidenceFactExact time
Action masking sets the probabilities of illegal actions to zero and renormalizes when the policy network outputs action probabilities, ensuring that the agent samples only from legal actions; it can significantly accelerate training but may introduce implicit preferences leading to path dependence.
1
Sources
50%
Confidence
Long-term
Relevance
10/29/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedAgents dynamically decide their next action based on context, and their behavior paths are generated only at runtime, breaking the traditional security model's assumption that the behavior of an executing entity can be enumerated.73% similarUnverifiedScheduling mechanisms establish clear temporal boundaries for proactivity, avoiding the trust and security risks posed by the uncontrollable random triggering of fully autonomous agents.71% 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.70% 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.69% similarUnverifiedThe Abliteration technique identifies and suppresses the direction vector responsible for refusal behavior in the model's residual stream, using orthogonal projection to remove that direction from the weight matrices, thereby eliminating refusal behavior without retraining.68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1112000API
curl https://kongchang.com/api/v1/knowledge/claims/1112000MCP
get_claim(id=1112000)