High-signal AI/security/automation notes.
IBM’s 602-organization breach study links AI-enabled attacks to $6 million average costs, while access controls, connected applications, cloud configuration, and machine identities remain the actionable failures.
UK AISI and NIST CAISI found Kimi K3 below leading frontier models in preliminary cyber tests, yet it completed a 32-step attack range once in ten attempts and did not refuse offensive tasks.
Five newly published CVEs affect the official MCP Ruby SDK through 0.22.0. Version 0.23.0 adds DNS-rebinding defenses and resource limits, but deployers must configure session ownership validation.
Proofpoint observed underground sellers advertising indirect prompt-injection generators for email, PDFs, calendar invites, and webpages. The activity is experimental, but it turns agent input channels into a concrete defensive priority.
CVE-2026-47427 crashes GitHub MCP Server versions before 1.1.0 with a malformed completion request. The patch is clear; the remote threat model needs nuance.
HashiCorp disclosed three Terraform MCP Server flaws spanning unauthenticated bearer-token theft, stateful session takeover, and stateless cross-tenant credential reuse. Version 1.1.0 fixes all three.
Netskope telemetry shows remote MCP transactions rising 375% in ten weeks while weekly downstream AI data-policy violations increased from 12 to 31 over a year. The report makes output authorization a first-class agent control.
Ruflo before 3.16.3 exposed an unauthenticated MCP bridge in its default Docker stack. The fix closes the shell path, but operators must investigate keys, MongoDB, and AgentDB memory.
Microsoft’s Project Perception coordinates red, blue, and green security agents, while MAI-Cyber-1-Flash routes most MDASH work to a specialist model. The design is notable; its launch benchmarks need careful reading.
n8n published 16 security advisories spanning host command execution, credential theft, MCP SSRF, file access, SQL injection, and cross-user isolation. Version 2.32.1 is the safest common remediation floor.
GuardianAgentBench tested 580 scenarios across three production agent frameworks. The best setup scored 74.8%, while runtime tool-call guardrails recovered 19.9% of failures at a 0.5% false-positive rate.
NVIDIA and 36 partners launched the Open Secure AI Alliance around an open agent-defense stack. The announcement names useful components, but no roadmap or governance model yet.
CVE-2026-55607 chained prompt injection, Git worktree confusion, symlinks, and fsmonitor to escape Claude Code’s macOS sandbox. Upgrade to 2.1.163 or later.
Anthropic’s Opus 5 system card reports strong exploit and cyber-range capability alongside materially lower prompt-injection success, with important limits for enterprise risk decisions.
CVE-2026-66012 exposed 31 SiYuan MCP tools through a Reader token, enabling workspace theft, credential exposure, file writes, and delayed desktop code execution. Upgrade to 3.7.2 or later.
UC Berkeley-led research separates untrusted exploration from privileged action and compresses cross-boundary hints, sharply reducing prompt-injection success without the utility collapse of rigid dual-agent baselines.
Accomplish chained CVE-2026-46331 through the local Claude Cowork Linux VM to reach a read-write share of the Mac host filesystem. Current Cowork sessions run remotely by default.
IssueTrojanBench reports 2,776 successful malicious actions across 4,176 coding-agent runs. Dependency installation was the most reliable attack; framework controls caused no measured rejections.
Redis shipped seven security releases after researchers published authenticated RCE chains against versions previously considered fixed. The Kimi K3 discovery-speed claims remain self-reported.
Eighteen Open WebUI CVEs disclosed on July 24 converge on version 0.10.0. Six are high severity, including cross-user tool execution and stored Pyodide XSS.