Desk · AI systems and model risk

AI Security

Coverage of AI application security, model abuse, agentic systems, and data exposure.

Articles: 63
Topics: 20
Page: 6 / 8

Data graphic: Plugin4Shell, pin bypassed. 2/4 agents patched: Claude Code patched in 2.1.179 and OpenAI Codex in 0.146.0; GitHub Copilot has no patch shipped and Gemini CLI won't be fixed.

Plugin4Shell: A Pinned Commit SHA Didn't Stop Repo Owners Swapping Plugin Code in Claude Code, Codex, Copilot and Gemini CLI

A branch named after a pinned commit SHA let plugin repo owners swap code in four AI coding agents. Claude Code and Codex are fixed; Copilot and Gemini CLI are not.

Oct 2 · 6 min

Data graphic: MCP server steals OAuth secrets. CVSS 7.5 High 6.5 for the interactive provider ; affected mcp 1.9.1–1.29.1 and 2.0.0–2.1.1, fixed in 1.30.0 / 2.2.0, M2M providers must also pass issuer=. An attack path runs from the MCP client to a malicious MCP server to the auth server it picked, which receives the client_secret and PKCE verifier.

highAI Security Model Context Protocol

Official MCP Python SDK Let Malicious Servers Steal OAuth Secrets, and Upgrading Isn't Enough

A flaw in the official MCP Python SDK let a hostile MCP server redirect OAuth client secrets. Fixed in 1.30.0 and 2.2.0, but M2M providers also need issuer=.

Oct 2 · 9 min

Data graphic: AI agent hacks the hunters. The attack chain from the internet through a Zammad session hijack CVE-2026-102489 to RCE, root CVE-2026-102490 and stolen volunteer data, with each flaw scored CVSS 4.0 9.4.

criticalAI Security Cve 2026 102489

AI Agent Breached Bug-Hunting Nonprofit DIVD by Chaining Two Zammad Zero-Days

An autonomous AI agent chained Zammad flaws CVE-2026-102489 and CVE-2026-102490 to reach root at DIVD. What Zammad admins should check and patch now.

Oct 2 · 8 min

Data graphic: TensorFold's local LLM server has 0 routes that check a credential: chat, completions, responses, models, health and metrics all run with no key. It is safe on the Mac default of 127.0.0.1:8080, reachable on every interface on the CUDA runbook's 0.0.0.0 path, and the two-GPU rendezvous on port 29551 is unauthenticated. Context: 10 releases from Sep 27 to 30, 175,000 exposed Ollama hosts, and no SECURITY.md.

watchAI Security Tensorfold

TensorFold Speeds Up Local LLMs but Leaves Its API Unauthenticated

TensorFold's fast OpenAI-compatible LLM server checks no credentials: safe on its localhost default, exposed on the 0.0.0.0 path its CUDA docs give.

Oct 1 · 6 min

ThreatFrontier data poster: GPT-6.1 Sol rated Critical for cyber at $2/$10 per million tokens, one-fifth of GPT-6 Astra's price, with 21.5% exploit success and 74.8% on DeepSWE v1.1.

elevatedAI Security Gpt 6 Sol

GPT-6.1 Sol Brings Critical-Rated Cyber Capability at a Fifth of Astra's Price

OpenAI's GPT-6.1 Sol nears GPT-6 Astra on coding at $2/$10 per million tokens, and inherits Astra's Critical cyber rating with a 4x jump in exploit success.

Sep 29 · 6 min

Dark briefing graphic comparing OpenClaw and Hermes Agent: 722 advisories vs 39 CVEs, top CVSS 9.9 vs 7.3, 341 malicious skills

elevatedAI Security Openclaw

OpenClaw vs Hermes Agent: Features, Local Models and the Security Record

How OpenClaw and Nous Research's Hermes Agent compare on architecture and local models, plus the major 2026 CVEs and a hardening checklist for both.

Sep 24 · 14 min

Data poster: Claude Opus 5.5, Anthropic's strongest cyber model yet, against Mythos 5.1 on ExploitGym 300 vs 264 exploits at six hours , control-flow hijacks 106 vs 81 and CyScenarioBench 67.6% vs 61.7% , with 73.4% full code execution on ExploitBench. A cyber classifier sends blocked requests to Opus 4.8.

elevatedAI Security Claude Opus 5 5

Claude Opus 5.5 Is Anthropic's Strongest Cyber Model Yet, So Most Security Work Goes to Opus 4.8

Opus 5.5 out-exploits Mythos 5.1 in Anthropic's own cyber tests and costs 40% less to run than Opus 5. Blocked security requests fall back to Opus 4.8.

Sep 22 · 15 min

Head-to-head data poster from SpaceXAI's Grok 4.7 launch table: Grok 4.7 xHigh against Claude Fable 5.1 Max on DeepSWE v1.1 71.0 to 70.0, EEBench 64.0 to 56.4, Harvey Legal 19.6 to 6.7, and Terminal-Bench 38.0 to 57.9, stamped VENDOR SCORED.

watchAI Security Xai Grok 4 7

Grok 4.7 Matches Claude Fable 5.1 Max — on SpaceXAI's Own Scorecard

SpaceXAI's launch table shows Grok 4.7 edging Claude Fable 5.1 Max on DeepSWE at a fraction of the price. On an independent harness, the gap is 29 points.

Sep 22 · 8 min