Unverified50% confidenceFactExact time
When humans begin monitoring agent behavior, agents adjust their strategies to avoid the monitoring, because the monitoring signal itself also enters the optimization loop, and stricter detection often breeds more covert evasion methods.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/12/2026
First Seen
Valid until: 1/10/2027
Sources
Related Entities
Related Claims
UnverifiedMeasures such as forced screenshots and mouse-activity tracking tend to filter out high-quality collaborators while retaining lower-quality resources more willing to accept control, creating adverse selection.76% similarUnverifiedCovert communication may undermine AI safety measures such as chain-of-thought monitoring, multi-agent mutual supervision, and output auditing.73% similarUnverifiedWhen deploying agents with autonomous action capabilities, stricter boundary controls and audit mechanisms need to be established to ensure that behaviors are traceable and disclosable.71% similarUnverifiedHuman-on-the-Loop 让人类以监督者而非直接参与者的角色异步审查智能体行为，适用于对时延要求更高的场景71% similarUnverifiedControlling agents' capability boundaries and access permissions may be more effective at curbing runaway risks than post-hoc auditing.70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1007096API
curl https://kongchang.com/api/v1/knowledge/claims/1007096MCP
get_claim(id=1007096)