{"slug":"security-injection","label":"Security Injection","item_count":3,"day_count":3,"source_count":2,"first_seen":"2026-10-16T11:10:30+00:00","last_updated":"2026-10-30T16:43:17+00:00","generated_at":"2026-11-03T00:12:58.466703+00:00","sources":["arxiv_agent_systems_research","arxiv_cs_ai"],"days":[{"date":"2026-10-16","items":[{"title":"MobileInject: Evaluating Reasoning-to-Action Security of MLLM-Based Mobile Agents under Cross-Modal Consistency Injection","url":"http://arxiv.org/abs/2610.23307v1","source":"arxiv_cs_ai","type":"paper","summary_1line":"Multimodal large language model (MLLM)-based mobile agents are rapidly evolving from conversational assistants into autonomous systems that perceive, reason over, and act on smartphone environments. This capability al...","why_it_matters":"Matches feed focus: agent, evaluation.","sid":"346a736923aff845","published":"2026-10-16T11:10:30+00:00","editor_note":"Evaluates reasoning-to-action security of MLLM mobile agents under cross-modal injection."}]},{"date":"2026-10-27","items":[{"title":"Heads of Security: Learned Attention Gates Improve Prompt Injection Robustness","url":"http://arxiv.org/abs/2610.37657v1","source":"arxiv_agent_systems_research","type":"paper","summary_1line":"We introduce Heads of Security (HoS), a prompt injection defense that enforces instruction-data separation via a small set of learned per-head attention gates. HoS learns how each head should redistribute attention be...","why_it_matters":"Matches feed focus: agentic, eval.","sid":"b483c5a93f031770","published":"2026-10-27T17:58:52+00:00","editor_note":"Introduces learned per-head attention gates as an instruction-data separation defense."}]},{"date":"2026-10-30","items":[{"title":"SecPO: Principled Adversarial Training for Prompt Injection Security","url":"http://arxiv.org/abs/2610.41998v1","source":"arxiv_cs_ai","type":"paper","summary_1line":"Recognized as a top threat to AI agents, prompt injection allows instructions from untrusted data (email, documents, websites) to manipulate the agent's operation. The existing best-defended LLM, Meta-SecAlign, is tra...","why_it_matters":"Matches feed focus: agentic.","sid":"b2f578a2e015ebec","published":"2026-10-30T16:43:17+00:00","editor_note":"Adds adversarial training aimed at beating Meta-SecAlign, the current best-defended LLM."}]}],"editorial":{"tldr":"MobileInject (Oct 16) framed prompt injection as a reasoning-to-action risk for MLLM-based mobile agents. Heads of Security (Oct 27) then proposed a defense: learned per-head attention gates that enforce instruction-data separation.","stale":false,"whats_new":"SecPO (Oct 30) proposes adversarial training against prompt injection, positioned against Meta-SecAlign, the existing best-defended LLM.","why_it_matters":"Prompt injection is a top agent threat, and the research is moving from attack evaluation toward defenses at the model level (attention gating, adversarial training).","take_for_builders":"Treat these as research, not controls: keep least-privilege tool scopes and confirmation on agent actions, and rerun your own injection tests before relying on any model-level defense.","status":{"state":"Developing","tone":"rising","changed":"2026-10-30","detail":"Three arXiv papers move from an attack benchmark to two model-level defenses; none is yet validated in production."},"beats":[{"kicker":"THE ATTACK","tone":"launch","headline":"MobileInject evaluates injection against mobile agents","sids":["346a736923aff845"]},{"kicker":"DEFENSE","tone":"rising","headline":"Heads of Security gates attention to separate instructions from data","sids":["b483c5a93f031770"]},{"kicker":"NOW","tone":"now","headline":"SecPO applies principled adversarial training","sids":["b2f578a2e015ebec"]}],"open_questions":["Do HoS or SecPO results hold against adaptive attackers outside their own evaluations?","Will either defense ship in an open-weight model or an agent framework?"],"generated_at":"2026-11-03T00:20:00+00:00"}}