14 articles tagged
OpenAI split Daybreak into Blue and Red tiers on 7 August 2026. Red bills 3.1x Blue's input rate, carries a long-context surcharge that can never fire, and is in no regional processing list.
The Cyber Jailbreak Severity scale gives AI jailbreaks what CVSS gave software bugs in 2005: a number everyone argues about instead of arguing about words. Here is how to score one, and what your team does at each band.
Zafran Security's FaceHugger research showed three ways a crafted Hugging Face repo runs Python during from_pretrained even when trust_remote_code is False. Here is the patch path and the policy.
Microsoft's Project Perception brings red, blue, and green team agents and the MAI-Cyber-1-Flash model to public preview on August 3, 2026. A build-versus-buy guide to agentic SOCs, with benchmarks, SCU costs
In July 2026 Anthropic's Claude Mythos Preview halved a post-quantum signature candidate's security margin and sped a reduced AES attack 200-800x. No production system broke, but the direction of travel is the story
On July 21, 2026, Gemini 3.5 Flash Cyber found 55 unique V8 vulnerabilities to Claude Opus 4.6's 36, then shipped only to governments and trusted partners. What the gated launch signals, and what AppSec teams can use
On July 27, 2026, NVIDIA and 36 partners launched the Open Secure AI Alliance and open-sourced the NOOA agent framework. Most of the tooling predates the alliance. Here is what is new, and a decision framework
Google Cloud's GKE AI security blueprint (16 July 2026) sets a 3-layer, 3-phase model for securing AI and agent workloads on Kubernetes: Confidential Nodes, Model Armor, gVisor sandboxing and org-level policy.
Microsoft's July 2026 Patch Tuesday fixed a record 570 flaws, including two zero-days already under attack. AI-assisted discovery is why the number exploded, and why the old triage signals stopped working.
Enterprises are deploying AI agents fast, but 88% had an agent-related breach last year and only 21% have mature governance. eCorpIT's AI agent security service closes that gap with assessment, red teaming, guardrails
JADEPUFFER exploited a critical Langflow flaw (CVE-2025-3248) and used an LLM to chain reconnaissance, theft and database extortion with no human operator. A breakdown of the attack and a 2026 defense checklist
CVE-2025-3248 gave unauthenticated attackers code execution on Langflow, and it was not the last. How to harden Langflow, Flowise, and n8n: patching, network isolation, gateway auth, secrets, and sandboxing for AI
Unit 42 found AI accelerated attacks 4x in 2026, with the fastest intrusions reaching data in 72 minutes and identity behind nearly 90% of breaches. Here is what changed and how CTOs and security leaders should respond.
Gartner expects over 40% of agentic AI projects to be canceled by 2027, and production agents succeed about 56.6% of the time. Three engineering lessons from shipping enterprise AI agents in 2026.
