This paper studies stealth memory injection, where a remote black-box adversary can use a single email payload to make persistent personal agents save poisoned memory while keeping the user-facing response quiet. It introduces WhisperBench, a 108-case full-cycle benchmark, and MemGhost, a one-shot payload generation framework for testing how poisoned memory later steers agent behavior.

VideoSEAL asks why long-video agents get more accurate without getting more grounded: a reward-hacking failure in multi-turn agentic RL. Under outcome-only GRPO, training-set accuracy improves while evidence-seeking behavior does not, because credit can flow to answer shortcuts rather than exploration actions. We trace this to two pressures: (1) reward pressure during training, where outcome-only rewards encourage the agent to speculate from insufficient evidence rather than reinforce evidence-seeking actions; and (2) prompt pressure at inference, where longer search traces saturate context and push planners toward speculative commitment instead of verification. The fix decouples an exploring planner from a frozen inspector with answer authority and abstention, enabling search-budget scaling and reducing semantic hallucination from 62.1% to 11.3% on LVBench.

This work shows that persistent personal agents like OpenClaw can suffer unintended memory pollution even without prompt injection: both user-attended foreground tasks and unattended background tasks may absorb ordinary external content into persistent memory. Because user-facing conversations and noisy tool-call results share the same session context, such content can lose provenance, be saved into long-term memory even without clear user awareness, and later steer user-facing behavior.

This work studies how to train clean models when both pre-trained models and fine-tuning datasets may contain unknown backdoor poisoning.

This work proposes an activation-guided framework to generate transferable prompt injection attacks against LLMs using gradient-free optimization.

This work investigates why mildly robust models generate more transferable adversarial examples than both naturally trained and highly robust models.

This work proposes a backdoor detection method based on topological evolution dynamics that is effective against both traditional and advanced backdoor attacks.

This work proposes a dynamic maximin optimization framework to improve the generalization of universal adversarial perturbations across models and samples.