Opening
Mu Yan starts with one thing today: a leaked API key from an AI inference provider ended up reaching all the way to a tampered Claude Code extension on users' computers. Today is Saturday, November 14, 2026, and we'll also talk about 186 knockoff domains impersonating well-known AI brands and what tricks they're pulling, plus how a forked AI penetration testing tool mapped out an entire AWS environment in 30 minutes. More on that in a moment.
Today's AI Security Highlights
1. Fireworks AI Key Leak Triggers Chain Reaction: Notion Chat Logs Exposed, Cursor Extension Tampered
- What happened: AI inference provider Fireworks AI confirmed that the scope of its breach has widened: billing data for all customers was affected, and some customers' API keys, datasets, and unfinished model weights were exposed. Even Notion's AI chat logs, covered by a zero-retention agreement, were caught up in the leak due to a backup mechanism (reportedly affecting under 1% of workspaces). Cursor further disclosed that the root cause of the October 29, 2026 incident, in which its Claude Code extension was swapped for a tampered version (2.1.312/2.1.313), was that attackers used these leaked Fireworks AI keys to steal internal data. Up to 5,000 users installed the compromised version.
- Who's affected: Enterprises using Fireworks AI's inference service (including Notion) and anyone who updated the Claude Code extension in Cursor around October 29, 2026 are within scope. In effect, a single leaked key rippled through several layers of the downstream supply chain.
- Defense recommendations: If your team uses Fireworks AI or other third-party inference services, start by inventorying which keys and datasets have passed through that platform and rotate them all. Cursor users should use the official cleanup script to check and reissue credentials, and always verify the version number after updating the extension going forward.
- Source: VentureBeat (https://venturebeat.com/security/attackers-breached-the-fireworks-ai-storage-where-notion-ai-chats-sat-outside-zero-data-retention), The Hacker News (https://thehackernews.com/2026/11/stolen-fireworks-ai-key-led-to-cursor.html)
2. AI Pentesting Framework Fork "AntiFox" Spotted in Real Attacks, Mapped an AWS Environment in 30 Minutes
- What happened: Researchers tracked "AntiFox," a fork of the AI penetration testing framework ARTEX, appearing in real-world intrusions, not just test environments. After attackers obtained a valid AWS identity, AntiFox fired off a flood of queries against more than a dozen services, including IAM, Secrets Manager, and Amazon Bedrock, within about 30 minutes, ultimately succeeding in reading out a production key.
- Who's affected: Any team that relies on long-lived IAM keys as its primary cloud identity credential is in scope. Once an identity or credential is stolen, an AI agent can map out an entire account far faster than a human attacker could.
- Defense recommendations: Replace long-lived IAM keys with role-based, auto-expiring temporary credentials wherever possible, and enable call-logging monitoring for Bedrock and Secrets Manager so abnormally high-frequency queries trigger real-time alerts.
- Source: GitGuardian (https://blog.gitguardian.com/another-new-ai-hacking-agent-vs-the-same-credential-problem-antifox-targeting-aws-credentials/)
3. "TypoPrompting": 186 Fake AI Tool Domains Trick Users Into Infecting Themselves
- What happened: Security firm Island exposed a campaign dubbed "TypoPrompting" (using typo-squatted domains that closely mimic well-known AI brand names to lure victims), uncovering 186 fake domains impersonating 27 AI brands including ChatGPT, Claude, and Gemini. These domains trick users into running ClickFix-style commands (getting users to paste a command into their own system and execute it), which then download browser-hijacking malware or install browser extensions that hijack search results. Installs of the extension pushed by a domain impersonating Perplexity alone have climbed to nearly 970,000.
- Who's affected: Any regular user who searches for AI tool names and might accidentally click into a fake site is a target. The risk is even higher at companies that don't control which browser extensions employees are allowed to install.
- Defense recommendations: Always download AI tools from official domains or app stores, and close any webpage that asks you to paste a command into a terminal or run dialog. Enterprises should consider using browser management policies to restrict where extensions can come from.
- Source: The Hacker News (https://thehackernews.com/2026/11/186-fake-ai-domains-redirect-users-to.html)
4. Anthropic's AI Scanning Tool First to Catch a Critical Apache HTTP/2 Vulnerability, CVSS 9.8
- What happened: Anthropic's OSS Scanner (an open-source code security scanning tool) was the first to flag a race condition vulnerability in the Apache HTTP Server mod_http2 module, which a Google Cloud researcher independently reproduced. Tracked as CVE-2026-119618 (CVE being the standard numbering system for publicly disclosed vulnerabilities), it carries a CVSS score (vulnerability severity rating, out of 10) of 9.8. Apache HTTP Server versions 2.4.55 through 2.4.70 are affected, letting an unauthenticated attacker execute arbitrary code remotely. Apache released a fix in version 2.4.71 on November 11, 2026, and says no in-the-wild exploitation has been observed so far.
- Who's affected: Any server still running Apache HTTP Server 2.4.55 through 2.4.70 with HTTP/2 enabled is at risk. These servers are commonly found as website front ends and reverse proxies, and a successful compromise hands the attacker remote code execution.
- Defense recommendations: Upgrade to Apache HTTP Server 2.4.71 as soon as possible. If an immediate upgrade isn't feasible, check whether mod_http2 is actually in use and consider temporarily disabling it to reduce exposure.
- Source: SecurityWeek (https://www.securityweek.com/apache-patches-critical-http-server-rce-found-by-anthropic-google/)
5. IBM Discloses Multiple Security Flaws in MCP Gateway Commonly Used by AI Agents, Upstream Credentials Could Be Overwritten via Header
- What happened: IBM issued multiple security advisories for ContextForge MCP Gateway (MCP, the standard protocol that lets AI agents connect to external tools), commonly used by AI agents, with versions 1.0.0 through 1.1.0 affected. One is an authorization bypass issue (CVE-2026-110195), and another is a flaw where upstream credentials can be directly overwritten via the X-Upstream-Authorization header (CVE-2026-110279, CVSS 8.1). IBM officially recommends upgrading to 1.1.0 and reviewing your upstream credential management configuration.
- Who's affected: Any team using ContextForge MCP Gateway to connect multiple upstream services and let AI agents call external tools on users' behalf should check this. Misconfigured credential management effectively lets someone else impersonate the configured identity when calling upstream services.
- Defense recommendations: Upgrade to ContextForge MCP Gateway 1.1.0, and review which headers are currently allowed through, making sure credential-related headers like X-Upstream-Authorization can't be set by the client.
- Source: IBM Security Bulletin (https://www.ibm.com/support/pages/node/7296489)
6. AI Takes the Front Line in Defense: OpenAI Launches Daybreak, Forescout Cuts Vulnerability Discovery Time From 9 Months to 90 Hours
- What happened: At its "Cyber at the Frontier" event on November 12, 2026, OpenAI launched two tools: Daybreak Red Team, which uses multiple agents to run penetration testing and attack path validation, and Daybreak Defense Factory, which can import vulnerability data reported by Wiz, Snyk, and Endor Labs to help generate remediation recommendations. OpenAI also announced a $1 billion commitment to subsidize access to these frontier security capabilities for critical service providers. The same day, Forescout's Vedere Labs announced a research framework built on off-the-shelf AI models like Claude Sonnet 4.6 and Claude Opus, which analyzed 10 common TCP/IP protocol stacks and verified 99 vulnerabilities (66 previously unknown, more than half rated high or critical severity), compressing roughly 9 months of manual analysis time down to 90 hours. 54 of those have already been reported to vendors for patching.
- Who's affected: Enterprises using vulnerability scanning tools like Wiz, Snyk, or Endor Labs may soon be able to turn scan results into fixes faster. Critical infrastructure vendors running on any of those 10 common TCP/IP protocol stacks should also check whether the 99 vulnerabilities Forescout found involve components they use.
- Defense recommendations: If your team is already using Wiz, Snyk, or Endor Labs, keep an eye on new tools that can connect to automated remediation suggestions. Teams running industrial control systems or IoT devices should check with vendors whether their equipment is on Forescout's disclosure list from this round.
- Source: SecurityWeek (https://www.securityweek.com/openai-unveils-daybreak-red-team-defense-factory-security-tools/), ComputerWeekly (https://www.computerweekly.com/news/366654735/Bug-hunting-time-cut-dramatically-with-AI-says-Forescout)
Closing
The one thing worth remembering today is that so many security incidents trace back to the same starting point: a leaked key. Whether it's the chain reaction from Fireworks AI or AntiFox's 30-minute sweep through an AWS environment, the opening move is always getting hold of a legitimate identity. Before you clock out today, take a few minutes to inventory which long-lived, never-rotated keys your team still has lying around. See you tomorrow.


























Comments