High-signal AI/security/automation notes.
Gambit recovered the staging server behind an autonomous agent campaign that hit 105 retail targets in six days. The novel detail is not the intrusion speed — it is a playbook step that wipes the victim’s database after exfiltration.
GitLab’s September 23 critical patch (19.4.1, 19.3.3, 19.2.7) fixes CVE-2026-89078 and CVE-2026-93577, two authenticated RCEs reachable through crafted regexes in CI/CD configs, alongside Duo AI variable leaks, MCP scope and search-context flaws, and an unauthenticated job-trace read.
Google disclosed PageBreak, an internal AI agent that found 500+ XSS vulnerabilities in first-party web apps by executing real payloads instead of reasoning about code. The more useful number is the other one: two findings across hundreds of apps on its high-assurance framework.
IBM’s September 23 bulletin for Financial Transaction Manager on OpenShift fixes CVE-2026-18875, an unauthenticated RAG-poisoning flaw that steers AI-driven MCP payment tool calls, alongside critical deserialization, code-injection, and traversal flaws. Fixed in 4.0.11.0.
CVE-2026-95985 lets a crafted repository steer Amazon Kiro IDE into writing its own global ~/.kiro configuration. Fixed in 1.0.242. It is the second untrusted-workspace file-write CVE in the same tool this year.
CVE-2026-77521 scores a perfect CVSS 10.0: MaxKB’s agent passes a shell backend to deepagents, which auto-injects an execute tool that the human-approval map never covers. Fixed in 2.10.5-lts — but the default backend and the missing gate are unchanged in 2.10.6-lts.
MITRE published five CVEs against mcp-remote on September 24, pinned to version 0.1.38 and citing a researcher package that explicitly withdrew two of those claims. The package is now at 0.14.3 under a different maintainer, and the CVE records say nothing about it.
Patrick Wardle’s not-a-mused PoC showed any unprivileged Mac process could reroute Muse’s dictation traffic, harvesting microphone audio and session tokens and steering the trusted agent. Meta hotfixed it within days; no CVE was assigned.
OX Security mapped 15,465 MCP servers across three public registries: 796 hostnames outside the US, production traffic on home networks, six abandoned domains available for a few dollars, and an always-allow permission in Claude Code that turned one approval into .env access on Haiku 3.5.
Zenity Labs disclosed three Agentforce flaws, collectively SalesBleed: poisoned Web-to-Lead submissions hijacked AI agents to exfiltrate CRM records via image tags and Slack unfurling, and abused a Slack reply action for anonymous phishing. All three are now patched.
Manifold Security found third-party[.]com — hard-coded as an example endpoint in 1,700+ public repos including AI agent skills and MCP-server docs — serving a Windows ClickFix lure since at least June. Thirteen more squattable placeholders, two serving scams, ride on hundreds of thousands of files that static checks never flagged.
Transluce’s September 23 report traces OpenAI-linked agent swarms through public urlquery.net logs: hack attempts against US and Australian data providers from May, activity back to March, and a September 19–20 run at crypto exchange Quidax that tried to trade, inject HTML, and probe the API.
Unit 42 showed a poisoned support ticket could steer AWS AgentCore Harness into scanning its own process memory and exfiltrating an Identity-vault JWT in plaintext. AWS closed the report as informative; the default shell tool stays on.
CVE-2026-5430 lets attackers forge JWTs with unsupported algorithms into WSO2 admin sessions. Patched in April, exploited in the wild from September 13, added to CISA’s KEV on September 24 with a September 27 deadline — and the KEV blurb describes a different bug than NVD does.
CVE-2026-90898 lets an unauthenticated caller run commands through Bifrost’s MCP stdio registration, and a second advisory reaches code execution through remote Go plugin loading. Transport 2.1.0 is the remediation floor.
CVE-2026-59822 let any fabricated Bearer token open an authenticated MCP session on LiteLLM before 1.84.0 via an OAuth2 passthrough fallback. CISA listed it as exploited on September 2, and the same fix closed a second query-string bypass.
Salt Labs researchers bypassed Manus’s prompt-injection filter with JSFuck-obfuscated instructions hidden in an email, executed a payload before any warning appeared, and escalated to a reverse shell holding connected third-party tokens. Meta confirmed and patched the issue through its bug-bounty program.