https://securitysonar.com/research/Recent content in Research on Security SonarHugoenThu, 08 Oct 2026 08:00:00 +0000https://securitysonar.com/research/open-weight-model-security-benchmarking-part-4/Thu, 08 Oct 2026 08:00:00 +0000https://securitysonar.com/research/open-weight-model-security-benchmarking-part-4/Part 3 showed that a harness’s own guardrail can’t see inside file content it reads. This installment layers a production-adopted detection standard (Agent Threat Rules) on top of the same attack data and finds something more useful than a pass/fail score: the same injected instruction, phrased two different ways, produced two completely different detection outcomes. One was caught by accident, one was caught correctly, and neither was caught the way you’d expect from reading the rule names alone.https://securitysonar.com/research/open-weight-model-security-benchmarking-part-3/Fri, 02 Oct 2026 09:00:00 +0000https://securitysonar.com/research/open-weight-model-security-benchmarking-part-3/Parts 1 and 2 benchmarked whether Qwen 3.8 is safe to deploy. This piece asks a different question: once that same model is wired into an agent with a shell and a filesystem, what stops it from doing something dangerous? An unmodified open-source harness, one variable toggled, 36 trials on the same local DGX Spark, with results that don’t split cleanly along “the model refused” or “the model complied.”https://securitysonar.com/research/open-weight-model-security-benchmarking-part-2/Mon, 28 Sep 2026 07:04:00 +0000https://securitysonar.com/research/open-weight-model-security-benchmarking-part-2/Part 1 made the case that security benchmarking is a discipline distinct from capability benchmarking. This installment puts it into practice with a fully local, air-gapped stack — NVIDIA DGX Spark serving Qwen 3.8-27B through Ollama, benchmarked with Meta’s CyberSecEval — and reports what actually happened, including a judge-model investigation that swung the same benchmark’s result from under 1% to over 70% malicious compliance depending on which model did the scoring.https://securitysonar.com/research/open-weight-model-security-benchmarking-part-1/Fri, 25 Sep 2026 07:30:00 +0000https://securitysonar.com/research/open-weight-model-security-benchmarking-part-1/Capability benchmarks and security benchmarks answer different questions, and most teams only run the first one. Using Alibaba’s newly released Qwen 3.8 as a case study, this piece lays out why open weights make independent security verification possible — and what the July 2026 ExploitGym sandbox escape proves about containment.https://securitysonar.com/research/the-harness-is-where-agent-security-lives/Wed, 23 Sep 2026 14:00:00 +0000https://securitysonar.com/research/the-harness-is-where-agent-security-lives/Three independent 2026 research efforts breached production AI agents without touching the model — and a parallel ‘harness engineering’ boom is building the exact same layer for reliability, largely without a security lens. The one teaching harness built with a security chapter outscores the popular shipped product that skipped it.https://securitysonar.com/research/how-transit-ai-handles-nondeterminism-in-network-operations/Tue, 15 Sep 2026 07:00:00 +0000https://securitysonar.com/research/how-transit-ai-handles-nondeterminism-in-network-operations/An LLM that proposes commands to a production router or firewall inherits the same run-to-run unpredictability as a chatbot, but the blast radius is different. A hands-on review of Transit AI’s network-ops agent shows what it looks like to make that unpredictability structurally irrelevant rather than filter it after the fact.https://securitysonar.com/research/cracking-passwords-at-ai-speed/Fri, 20 Mar 2026 12:00:00 +0000https://securitysonar.com/research/cracking-passwords-at-ai-speed/Apple’s mode 14800 backup KDF is a sequential, memory-latency-bound workload that discrete GPUs handle badly. On a DGX Spark, unified memory between the Arm CPU and Blackwell GPU cracked a 6-digit iOS backup PIN in under 24 minutes at 70W — less than half the time an RTX 4090 needs.https://securitysonar.com/research/ai-networking-rethink-network-security/Fri, 20 Feb 2026 12:00:00 +0000https://securitysonar.com/research/ai-networking-rethink-network-security/AI fabrics move the most critical traffic memory-to-memory, bypassing the CPU and the kernel entirely. Every network security tool that treats the kernel as its primary witness — Zeek, EDR, NetFlow analytics — goes blind to it. Here’s what changes and where visibility has to move.https://securitysonar.com/research/data-poisoning-witches-brew/Sun, 11 Jan 2026 15:00:00 +0000https://securitysonar.com/research/data-poisoning-witches-brew/Witches’ Brew is a clean-label poisoning attack: the malicious training images look like ordinary data to a human reviewer, but carry a gradient aligned to a hidden target. File-integrity checks pass. The model’s logic is subverted anyway.https://securitysonar.com/research/the-new-mandate-hands-on-agentic-ai/Sun, 11 Jan 2026 09:00:00 +0000https://securitysonar.com/research/the-new-mandate-hands-on-agentic-ai/The autonomous deployment of agentic AI is outpacing our ability to assess its risks. The most effective security professionals won’t be the ones waiting for perfect frameworks — they’ll be the ones who understand the technology by building it.