On November 16, Google’s Threat Intelligence Group and Mandiant disclosed exploit_agent, a multi-agent AI framework discovered on an exposed staging server that diffs a vendor’s patch against the prior version, identifies the underlying vulnerability, drafts a working exploit, and fires it, with no human directing any step. GTIG also confirmed two other autonomous frameworks, CAIRN and PentAGI, operating against live targets. This is not a research demo. CAIRN has already been tied to breaches at ten organizations in advanced manufacturing, defense, and civil-rights sectors. A sibling framework, ARTEX, has been linked to the recent Barracuda compromise and to Storm-3358 campaigns running since August.

Practitioners should care about this one specifically, not generically, because it attacks the assumption that patch cadence buys you time. Every vulnerability management program is built on the idea that there’s a gap, hours or days or weeks, between a patch dropping and a working exploit existing. Exploit_agent is purpose-built to close that gap without a person in the loop, which means the gap stops being a function of attacker skill and starts being a function of compute availability.

Analysis

The distinction GTIG draws between exploit_agent and the earlier-known ARTEX matters more than it sounds. ARTEX, the framework behind Barracuda and the Storm-3358 campaigns, is an autonomous penetration-testing tool: it automates reconnaissance through exfiltration, but a human still points it at a target and steers the engagement. Exploit_agent removes that operator from the specific pipeline that worries defenders most: turning a patch diff into a weapon. Patch diffing has always been a known technique for finding the vulnerability a vendor just fixed; what changes here is that the entire chain from diff to drafted exploit to launch runs unattended. Three frameworks doing variations of this at once, found active rather than theorized, is the real story. It is no longer one actor’s novel tool. It is becoming infrastructure.

Java4Shell, tracked as CVE-2026-114938, is the live test case for what this means in practice. The OpenJDK Base64 decoder flaw enables unauthenticated RCE against default Jira, Jenkins, Keycloak, and Confluence configurations on AVX-512-capable hardware. Oracle and the OpenJDK project patched it the same day Calif disclosed it, November 17, and CISA pushed an alert. As of November 19, GreyNoise has logged probing from a handful of IPs with no confirmed compromise. SecurityWeek reported that VulnCheck assesses real-world exploitation still requires per-target engineering work, which is the only reason this hasn’t already turned into a mass-exploitation event. That per-target friction is precisely the ceiling an autonomous patch-diffing agent is designed to climb. Java4Shell is a preview of what a Jira or Jenkins zero-day looks like once that friction is automated away.

SignDrop is the other half of the picture: a Windows privilege-escalation zero-day with public exploit code and no patch, defeating the CLFS authentication mitigations Microsoft shipped in October 2025 specifically to kill this bug class. When there’s no patch at all, the “time to exploit” clock defenders usually rely on reads zero from day one. Put SignDrop next to exploit_agent and the uncomfortable conclusion is that the no-patch case, long treated as the rare, worst-case scenario, is becoming the baseline case attackers should be assumed to reach for once autonomous tooling makes vulnerability discovery cheap independent of a patch artifact to diff against.

The second major development this week attacks a different trust anchor: code signing. OPTIMJACK compromised Optimizely’s web-experimentation platform, almost certainly through the account signup flow, and injected malicious JavaScript into 188 CDN-hosted snippets across 178 organizations, including Salesforce, Zoom, GitLab, HSBC, Equifax, Air France-KLM, Ledger, BBC, CNN, and Ikea. Because the code shipped signed with Optimizely’s own CDN certificate, browsers and site operators had no reason to flag it. For roughly 19 hours before Optimizely remediated on November 16, visitors to those sites received a crypto clipper, a card skimmer, and an iOS exploit loader. The campaign was broader than Optimizely alone: researchers traced the same Coruna and DarkSword exploit-kit infrastructure through at least three other script providers, Kissmetrics, StorifyMe, and minute-ly, feeding zero-click exploitation of Mobile Safari across at least 21 major sites. Code signing answers “did this come from where it claims to,” not “is this safe to execute.” Attackers only needed to compromise the signer once to make every downstream customer complicit without any of them doing anything wrong.

Both stories erode the same category of control: a signal defenders treat as a proxy for safety that attackers have learned to attack directly instead of routing around. Patch status and code signatures are provenance and timing checks, not behavior checks, and this week gave two separate demonstrations of why that distinction now matters operationally rather than academically.

Escalations from Monday

Monday’s report covered N-able shipping patches for 17 N-central zero-days, including bugs already being exploited to deploy crypto miners on customer on-premises servers. This week N-able disclosed a separate incident: unauthorized access to a server inside its own hosted Mail Assure and SpamExperts cloud, forcing the company to disable management interfaces and APIs across the Americas, Europe, and APAC. N-able says it has no evidence yet of mailbox or customer data access. The distinction matters: Monday’s story was about customers failing to patch fast enough. This one is about the vendor’s own cloud infrastructure getting breached, which every downstream MSP inherits regardless of their own patch discipline. For an RMM and email-security vendor, that’s a materially worse category of exposure than an unpatched CVE.

What to watch

Watch for exploit_agent-class tooling to surface against a specific named CVE before Monday’s report. GTIG’s disclosure describes capability found on an exposed server; confirmation that it produced a working exploit against a live target would move this from “capability exists” to “capability is operational,” and that’s the threshold that should change how fast organizations treat patch windows. Also watch whether SignDrop gets a Microsoft patch before it gets weaponized at scale; every day without one extends a zero-day window that already has public exploit code behind it.

Security Unlocked publishes threat intelligence and strategic analysis twice weekly. This mid-week brief covers developments from Monday, November 16, 2026 through Thursday, November 19, 2026.