{"slug":"attack-security","label":"Attack Security","item_count":3,"day_count":2,"source_count":2,"first_seen":"2026-11-10T11:52:40+00:00","last_updated":"2026-11-19T16:47:14+00:00","generated_at":"2026-11-20T00:13:34.686460+00:00","sources":["arxiv_agent_systems_research","hackernews_ai"],"days":[{"date":"2026-11-10","items":[{"title":"From Attack Surface to Defense Architecture: Compositional Security for Agentic Coding Harnesses","url":"http://arxiv.org/abs/2611.13960v1","source":"arxiv_agent_systems_research","type":"paper","summary_1line":"AI coding harnesses couple language models to shells, filesystems, and network sockets, turning the harness into an automated Living-off-the-Land (LotL) platform: once an adversary injects instructions into retrieved...","why_it_matters":"Matches feed focus: agentic, harness.","sid":"455c0c72bf13295c","published":"2026-11-10T11:52:40+00:00","editor_note":"Sets the threat model: tool-connected coding harnesses as automated Living-off-the-Land platforms after injection."}]},{"date":"2026-11-19","items":[{"title":"AI security engineering: the same failures, but new ways to attack","url":"https://zimmergren.net/ai-security-engineering","source":"hackernews_ai","type":"news","sid":"1b66b3acc2678fad","published":"2026-11-19T09:59:53+00:00","editor_note":"Practitioner view that AI security repeats familiar failure classes through new attack paths."},{"title":"Did the attack work? Auditing AI-agent security benchmark verdicts","url":"https://zenodo.org/records/24321025","source":"hackernews_ai","type":"news","summary_1line":"Did the attack work? Auditing AI-agent security benchmark verdicts","why_it_matters":"Matches feed focus: agent.","sid":"cd759e6684ea23fc","published":"2026-11-19T16:47:14+00:00","editor_note":"Moves the thread to measurement: audits whether benchmark attack-success verdicts are correct."}]}],"editorial":{"tldr":"This is a loose research thread on agent security, not one event. It opened with an arXiv paper arguing that coding harnesses wired to shells, filesystems, and sockets become Living-off-the-Land platforms once an injected instruction lands.","stale":false,"whats_new":"A Nov 19 audit questions whether AI-agent security benchmarks score attacks correctly, asking if reported \"attack succeeded\" verdicts actually hold up.","why_it_matters":"If benchmark verdicts on agent attacks are unreliable, the security scores you use to pick a harness or guardrail may not reflect real exposure.","take_for_builders":"Before trusting an agent-security benchmark score, spot-check a sample of its \"attack succeeded\" and \"attack failed\" verdicts against the raw transcripts.","beats":[{"kicker":"THREAT MODEL","tone":"launch","headline":"Coding harnesses framed as Living-off-the-Land attack platforms","summary":"Paper maps prompt injection through retrieved content to shell, file, and network actions, and proposes compositional defenses.","sids":["455c0c72bf13295c"]},{"kicker":"NOW","tone":"now","headline":"Practitioner take on old failures, plus an audit of benchmark verdicts","summary":"Attention shifts from attack surface to whether the measurements of attack success can be trusted.","sids":["1b66b3acc2678fad","cd759e6684ea23fc"]}],"open_questions":["Which named agent-security benchmarks does the verdict audit find mis-scoring attacks, and in which direction?","Do benchmark maintainers publish corrected verdicts or revised scoring rules?","Does the compositional defense architecture get an implementation or evaluation against a real harness?"],"generated_at":"2026-11-20T00:12:40Z"}}