Opening
On November 16, 2026, Google disclosed the first fully automated AI cyberattack: an AI framework called exploit_agent that found vulnerabilities and wrote exploit code entirely on its own, with no human involvement at any stage. Meanwhile, Amazon caught a self-replicating agent hunting for AI API keys, and a patched UnrealIRCd vulnerability was already being exploited in the wild the very next day. Coming up, I'll tell you how underground marketplaces are now using AI to write "product descriptions" for hacked computers. I'm Muyan, and this is AI Security Daily.
Today's AI Security Highlights
1. UnrealIRCd Patch Exploited the Very Next Day, CISA Adds It to Mandatory Patch List
- What happened: The UnrealIRCd maintenance team used AI to assist with an internal code audit and uncovered a command injection vulnerability tracked as CVE-2026-114713 (CVSS 8.6), one of the most severe issues found in the project in the past 15 years. The team released patch 6.2.7.1 on November 7, 2026, but the very next day, November 8, 2026, it was already being exploited in real-world attacks. CISA added it to the Known Exploited Vulnerabilities (KEV) catalog on November 14, 2026, requiring federal agencies to patch by November 16, 2026.
- Scope of impact: Any server still running an older version of UnrealIRCd is at risk, especially operations teams whose patching cadence can't keep up with the pace of disclosure.
- Defense recommendations: Upgrade to 6.2.7.1 or later immediately. If you can't upgrade right away, restrict external connections in the meantime and check for any signs of unusual access.
- Source: The Hacker News (https://thehackernews.com/2026/11/attackers-exploit-unrealircd-flaw-that.html)
2. Google Discloses exploit_agent, the First Fully Automated AI Cyberattack
- What happened: In its Q4 AI threat report released on November 16, 2026, Google's Threat Intelligence Group (GTIG) documented the first fully automated attack case on record. According to the report, suspected Russian-speaking hackers built a framework called exploit_agent that automatically diffs pre- and post-patch code, identifies vulnerabilities, and generates exploit code, all without any human intervention. Reported targets included media organizations, educational institutions, and aerospace entities in France.
- Scope of impact: In theory, any software project that publicly releases patches could become training ground for this kind of automated diffing framework, with media, education, and aerospace organizations facing higher risk due to their relatively limited patching resources.
- Defense recommendations: Speed up internal validation and deployment as soon as a patch is released, and minimize the gap between a patch going live and your own environment getting updated.
- Source: The Register (https://www.theregister.com/research/2026/11/16/google-ai-agent-turned-a-security-patch-into-a-working-exploit-with-no-human-help/5309926)
3. Amazon Catches Self-Replicating AI Attack Agents Targeting AI App Keys
- What happened: Since October 3, 2026, Amazon's honeypot system MadPot has repeatedly caught AI attack agents scanning AI applications like LiteLLM, One API, and Dify, exploiting weaknesses such as default admin keys to steal credentials belonging to upstream AI providers. The stolen keys are then used to power even more agents that go hunting for the next batch of keys, forming a self-replicating attack chain. Amazon assessed with medium confidence that the operation uses a framework called ARTEX and may be linked to a China-affiliated group.
- Scope of impact: Any team running AI application gateways like LiteLLM, One API, or Dify with admin keys still set to defaults is a potential target. Once those keys leak, attackers effectively gain the quota and access rights to your upstream AI services too.
- Defense recommendations: Check whether the admin interfaces for these AI applications are still using default keys, replace them with dedicated, least-privilege keys, and set up alerts for abnormally high volumes of API calls.
- Source: The Hacker News (https://thehackernews.com/2026/11/amazon-uncovers-ai-agent-campaign.html)
4. Underground Marketplaces Use AI to Write "Product Descriptions" for Hacked Computers
- What happened: On November 14, 2026, Hudson Rock's Infostealers research team discovered a new underground marketplace on a crime forum that uses AI to automatically generate searchable "product descriptions" from the logged data of each stolen computer. Buyers can search for a specific company's VPN or cloud access using natural-language queries like "Citrix access at a healthcare organization," with bids starting as low as $0.25.
- Scope of impact: Any corporate device that has ever been infected with infostealer-type malware could end up packaged into an easily searchable, easily biddable product, lowering the bar for attackers looking to target a specific company.
- Defense recommendations: Treat employee credential leaks as a given risk, enforce multi-factor authentication across the board, and regularly check whether your company domain shows up in leaked-credential records.
- Source: Cybersecurity News (https://cybersecuritynews.com/hackers-use-ai-powered-marketplace)
5. Palo Alto Networks Report: 99% of Enterprises Have Shadow AI They Can't Keep Under Control
- What happened: A new Palo Alto Networks report found that over the past year, 99% of surveyed enterprises discovered unmanaged "shadow AI" in their environments, yet only 9% could automatically inventory and classify it. 73% of enterprises allow AI agents to modify code or production configurations, 40% can't reconstruct an agent's actions after the fact, and it takes an average of 2.5 days to revoke a compromised non-human identity credential.
- Scope of impact: This applies to almost every enterprise using AI agents in development or operations workflows, with the highest risk falling on teams that lack a unified inventory of agent permissions and have no dedicated identity management mechanism.
- Defense recommendations: Give every AI agent its own dedicated identity with least-privilege access, and build a process that can revoke credentials quickly, cutting revocation time from days down to hours.
- Source: SiliconANGLE (https://siliconangle.com/2026/11/16/palo-alto-networks-report-finds-organizations-struggle-to-track-and-secure-ai-agents-in-the-cloud/)
6. Defensive Cybersecurity AI Startup Corma Closes $125M Series A
- What happened: Defensive cybersecurity AI startup Corma announced it closed a 1 billion. The company has now completed two funding rounds in just three months since its founding. Corma trains its own proprietary defensive models and AI agents to help enterprise security teams with threat investigation and incident response.
- Scope of impact: This is an industry development reflecting a broad consensus in the security community: to keep up with the kind of automated attacks described above, defenders also need AI agents to speed up their response. It's not an immediate risk to any single company.
- Defense recommendations: Before evaluating AI-assisted security tools, make sure your own alerting and incident-response records are thorough enough: that's what gives the tool something to learn from.
- Source: Unite.AI (https://www.unite.ai/corma-raises-125m-series-a-at-1b-valuation-to-scale-defensive-ai/)
Closing
The one thing worth remembering today is that attacks can now go from finding a vulnerability to writing exploit code entirely without human hands, which means defenders can no longer pace their response to human speed. If your AI applications are still using default admin keys, or nobody can tell you exactly how many AI agents are running across your company, put it on today's checklist. I'm Muyan, see you tomorrow.


























Comments