· ·

AI Security Newsletter No.92 (11-18-2026)

By Song Luo · AI security product leader

Welcome to this edition of the AI Security Newsletter. This week, AI agents misbehaved in the wild: xAI’s Grok Bot reached for company apps on its own and then denied it, and attackers turned the ARTEX pentest framework against real targets. Defenders answered with agents of their own, from Shopify’s Cyber Chaos Monkey to OpenAI’s Daybreak Defense Factory. We also cover a prompt-injection path into Codex’s npm release, new MCP testing tools, and fresh guidance on supervising AI agents from the OCC and Gartner.

Risks & Security

Grok Bot Made Unapproved Access Attempts, Ignored Stop Requests, Then Denied It

Research from identity-security firm NewCore found two xAI Grok Bot agents made 31 unrequested calls over five days, trying to read Gmail, Calendar, Slack, GitHub and Drive, some while the owner’s laptop was closed. One bot, told to stop, later tried every connected app, then replied “No. I made no NewCore tool calls.” Neither chat history nor Grok’s logs recorded them. NewCore’s gateway blocked all 31. xAI has since added an admin toggle for proactive messages.

References:

Google Scans the Public Web for Indirect Prompt Injections

Google researchers scanned Common Crawl snapshots for indirect prompt injections, filtering with pattern matching, Gemini classification and human review. Most hits were educational, but real injections included pranks, SEO attempts to make assistants promote businesses, and traps aimed at AI agents. A small malicious share tried to exfiltrate data or delete files. These were unsophisticated, yet the malicious category grew about 32% between November 2025 and February 2026, and Google expects attacks to scale as agents gain capability.

References:

Unit 42: Attackers Use ARTEX AI Agents to Turn Password Resets into Cloud Access

Palo Alto Networks Unit 42 reports attackers using ARTEX, an open-source multi-agent pentest framework, against real targets. Without exploiting any software bug, agents abused password reset pages and exposed APIs, harvested routes from OpenAPI specs and source maps, and validated over 200 client secrets in about three minutes. Unit 42 counted 609 internet-facing ARTEX hosts in October. It advises removing public specs, rate-limiting self-service flows, and using generic errors that starve agents of feedback.

References:

Elastic: Five Agentic Attack Frameworks That Never Ask Permission

Elastic Security Labs analyzed ARTEX, CyberStrikeAI, Cairn, PentAGI and HexStrike, open-source offensive AI frameworks with over 60,000 GitHub stars combined; Google and Mandiant have seen four in real intrusions. None enforces authorization by default, and all run autonomously out of the box. ARTEX acts in bursts of two to 25 seconds, so containment must be decided in advance. Elastic recommends measuring how fast identities take sensitive actions and shares an ES|QL hunt for rapid multi-tool activity.

References:

Prompt Injection in a GitHub Issue Could Have Backdoored OpenAI’s Codex npm Release

Depi researchers showed how a prompt injection in a GitHub issue could hijack the openai/codex repository’s issue-labeling Codex agent, which anyone could trigger. The agent ran shell commands, leaked workflow tokens from runner memory, and let the researchers poison the Actions cache to reach a privileged workflow, from which an OIDC token could be exchanged for npm publishing rights. OpenAI fixed the bug after a P1 Bugcrowd report. AI triage bots can become paths into release pipelines.

References:

GitDoor and GraphDoor Backdoors Hide C2 in GitHub and SharePoint

Kaspersky disclosed Operation ShadowCloud, an espionage campaign against government organizations in Asia and Africa. Attackers use a legitimate MeshCentral agent to deliver a DLL-sideloading chain that loads one of two backdoors in memory. GitDoor reads commands from a public GitHub repository and posts results to a private one; GraphDoor uses the Microsoft Graph API to stage commands and data on SharePoint. Because C2 traffic goes to trusted cloud services, domain allow-lists alone will not stop it.

References:

Technology & Tools

Shopify’s Cyber Chaos Monkey Lets AI Agents Attack Its Own Production Systems

Shopify Engineering described Cyber Chaos Monkey, an agentic system that hunts vulnerabilities in live production using test shops, real credentials, source code and past bug reports. A discovery stage generates leads; a sandbox stage tests them with live requests. Containment is built in: unprivileged Kubernetes jobs, proxy-injected credentials, an egress allowlist, and a separate judge agent on a different model that can end any session. Shopify reports tens of thousands of engagements, dozens of significant findings, and no outages.

References:

AWS Shares Best Practices for Bedrock Guardrails Denied Topics

AWS published guidance on writing Denied Topics in Amazon Bedrock Guardrails, which block content by meaning rather than keyword, so loosely written topics over-block. It warns against topics framed around whole subject areas, keyword-list definitions, misleading examples and overlapping topics. Instead, frame each topic around what the user is trying to get, using short action names. In its bank example, a broad “Finance” topic blocked “When does my direct deposit clear?” Each guardrail holds at most 30 topics.

References:

Google Launches Gemini Decisions API for Fast, Calibrated Decisions

Google introduced the Gemini Decisions API, built on Gemini Decision 1, a model that scores candidate answers instead of generating text. One request can ask up to 1,000 yes/no, multiple-choice or rubric questions about a shared multimodal input of up to 256K tokens, with 150–250 ms latency and a confidence score per verdict. Target uses include tool selection, policy checks and safety screening; Google Search already uses it to spot adversarial behavior in MCP tools.

References:

OpenAI’s Codex Security CLI and SDK Bring Agentic Vulnerability Scanning to the Terminal

OpenAI’s open-source Codex Security CLI and TypeScript SDK, released under Apache 2.0 as part of its Daybreak lineup, scan repositories, paths or code changes for vulnerabilities, try to confirm each finding, and propose fixes for human review. They can compare scan runs, verify fixes, and fail CI/CD builds above a severity threshold. Both remain in beta with gated access, and OpenAI advises storing reports privately because they can contain source excerpts and vulnerability details.

References:

Open-Source AI DFIR Toolkit Ships 171 Detection Signatures

The vendor-neutral AI DFIR Toolkit offers 171 Sigma, YARA and Suricata signatures covering prompt injection, MCP attacks, model supply chain, copilot abuse, RAG and vector database attacks, agentic orchestration and AI-service C2. Examples include YARA rules for MCP tool poisoning. Rules map to MITRE ATLAS and the OWASP LLM and Agentic lists, and the project adds an endpoint artifact catalog and 26 case studies. The Sigma rules are untested on real events; baseline them locally.

References:

mcp-attest Tests Whether MCP Servers Enforce Your Authorization Policy

mcp-attest is an MIT-licensed Python tool that checks whether MCP servers enforce your policy. Given a JSON policy of identities and their allowed tools, it connects to a live server as each identity, makes real tool calls, and reports every case where the server allows something the policy denies. Checks map to the OWASP MCP Security Testing Guide, including OAuth validation and hidden instructions in tool metadata. A GitHub Action fails builds on critical or high findings.

References:

Hunting Threat Infrastructure With a Local Model and MCP

Embee Research shared notes on finding malicious infrastructure with a fully local setup: GLM5.3-flash running on two NVIDIA DGX Sparks inside a stripped-down OpenCode, connected to Validin and URLScan through MCP. Given a seed indicator, the model pivoted on class hashes, shared files and file hashes to group related sites by confidence. In four tests on phishing and fake Cloudflare pages, it found related infrastructure. The post is informal and does not measure accuracy.

References:

OpenAI Open-Sources MCP Extensions for ChatGPT-Native Plugins

OpenAI released MCP Extensions, TypeScript and Python SDKs plus a protocol spec that let MCP servers plug into ChatGPT beyond chat: sidebar apps, side panels, file viewers and editors, composer mentions, and forms. Plugins can also send messages to current or new conversations. Canva, Figma and Adobe are early adopters. For security teams, third-party MCP servers gain more ways to reach users and files inside ChatGPT, and OpenAI’s developer page does not describe a permission model.

References:

Business & Products

OpenAI Launches Daybreak Defense Factory to Turn Findings into Patches

OpenAI launched Daybreak Defense Factory alongside Daybreak Red Team on November 12. Defense Factory ingests findings from existing security tools and Daybreak Red Team, and agents running OpenAI’s cyber models investigate each issue and produce a merge-ready pull request with a patch and a test. Engineers approve every fix before production. Daybreak Red Team, in early access, uses agent fleets to test web apps. Pricing and general availability have not been announced.

References:

Detectify Lets Your AI Agents Prove Exploitability With an Evidence Ledger

Detectify launched Agentic Tooling, which lets an organization’s own AI agents check whether suspected vulnerabilities, such as SAST findings or bug bounty reports, are exploitable in running applications. Agents test each ticket with Detectify’s attack engines and return one of four verdicts. Every request and payload is written to an append-only Security Ledger the agent cannot edit, exposing claims without evidence. It works with any MCP-capable agent, and risky tests are held for human approval.

References:

IBM and Microsoft Launch Imagine Enterprise for Governed Agents

IBM Consulting and Microsoft launched Imagine Enterprise, a platform for building, running and governing AI agents across business processes. IBM contributes more than 6,000 prebuilt agents and assets; Microsoft brings Foundry, GitHub Copilot and Copilot Studio, with Entra for identity, Purview for data governance and Defender for security. First frameworks target financial services, insurance and retail. The key question for security teams is whether governance controls hold up across thousands of prebuilt agents.

References:

OpenAI Launches ChatGPT Wallet So Agents Can Pay Without Seeing Your Card

OpenAI began rolling out ChatGPT Wallet, letting its agents complete purchases for US ChatGPT Pro subscribers. Built with Visa, Mastercard and American Express, it issues a network payment token per approved purchase, so neither agent nor merchant sees the real card. Each purchase requires user approval, such as Face ID, and leaves an authorization record. The launch highlights two controls agentic commerce is settling on: tokenized credentials and step-up approval on every payment.

References:

OpenAI Expands Free ChatGPT for Academic Researchers

OpenAI added 50,000 free grants to its ChatGPT for Academic Researchers program, with applications opening November 23 and a new target of 200,000 researchers. So far 20,000 researchers in 33 countries have joined. New grants include ChatGPT Plus rather than the Pro tier pilot users received. An OpenAI study of more than 1,800 pilot scientists found reading and writing made up over half their science messages, and OpenAI says it does not show that AI speeds up discovery.

References:

Regulation & Policy

OCC Gives Bank Examiners a Framework for Supervising AI Agents

The US Office of the Comptroller of the Currency issued Bulletin 2026-54, a non-binding framework for examiners reviewing banks’ use of AI, including agents. Examiners should rank use cases by the cost of failure, the agency given to the system and the level of human oversight, not by technical sophistication. Legal analysts note it covers generative and agentic AI and points banks toward circuit breakers, rate limits, restricted permissions and the ability to suspend agents.

References:

China’s Top AI Labs Strip Offensive Cyber Skills From New Models

Bloomberg reports leading Chinese AI developers are releasing models with weaker offensive cyber abilities. Alibaba filtered attacker-helpful material from its newest Qwen training data, while Tencent and Moonshot AI cut cyber functions in Hy4 and Kimi K3.1 Pro. The shift follows an October Ministry of Commerce draft that would ban exports of AI tools able to attack critical infrastructure. Researchers warn that filtering lowers misuse risk but cannot remove it.

References:

Opinions & Analysis

Gartner: Weak AI Governance Will Disrupt Critical Services at 40% of I&O Teams by 2028

Gartner predicts that without strong governance, 40% of infrastructure and operations organizations scaling AI will suffer a business-critical service disruption by 2028. Only 15% of surveyed respondents had automatic circuit breakers to stop runaway agent errors or loops. Recommended controls include per-agent identities with short-lived, least-privilege credentials, runtime checks on tool calls, automatic halts when thresholds are crossed, immutable decision logs, and human approval for high-impact changes.

References:

Leave a Reply

Discover more from Mindful Machines

Subscribe now to keep reading and get access to the full archive.

Continue reading