On September 21, an AI agent broke into a cybersecurity organization, chained two zero-day vulnerabilities, and reached root on the host in seconds. No human was steering it. Eight days later, Anthropic published its assessment of GLM-5.3, an open-weight Chinese model that can build working exploits on its own and whose guardrails can be stripped for about $4,400. And two months before any of that, the Department of War paused CMMC Phase II.
These facts point to a larger issue that the defense industrial base cannot afford to ignore.
Nation-state and criminal attackers can now target more companies without adding skilled operators. Third-party CMMC assessments made cybersecurity a priority for Defense Industrial Base (DIB) leaders by tying it to contract awards. With assessments paused, that urgency has faded, even as threats grow faster and cheaper.
On July 13, 2026, the Department of War suspended CMMC Phase II, the stage that would have required third-party C3PAO assessments on most Level 2 contracts starting November 10. All pending and future CMMC milestones were frozen with it. A reform task force was stood up under the DoD CIO, and on September 3 a class deviation turned the pause into binding contract language that stays in force until DoD cancels it.
CIO Kirsten Davies cited SBA data estimating future CMMC phases could cost small and mid-sized businesses over $7 billion annually, emphasizing tangible cyber hygiene over costly certifications.
While the pause shifts incentives, key obligations remain:
What disappeared was the deadline that pushed cybersecurity onto boardroom agendas and made it a priority. In most small companies, a requirement without a date becomes a project without a budget.
The DIB is where the most sensitive unclassified design, test, and sustainment data lives, spread across tens of thousands of small firms. Adversaries know that the weakest link in a program is rarely the prime, and if thousands of those firms deprioritize the work they were doing to get ready for November, the base gets softer as the threat gets sharper.
For most of the history of cyber espionage, the limiting factor on a nation-state program was people. Finding a vulnerability, writing a working exploit, and running an intrusion took skilled operators, and there were only so many of them. That constraint is what kept most small defense suppliers off the target list.
GLM-5.3, released August 14 by Zhipu AI (Z.ai ), shows that constraint is gone. NIST identified it as the most cyber-capable open-weight model to date. Anthropic found it develops end-to-end exploits for real Chrome V8 vulnerabilities at nearly the same rate as Claude Mythos Preview, a model restricted to vetted defenders due to its capabilities. A year ago, all models scored near zero on that test.
In two Anthropic sessions, AI models demonstrated advanced exploitation capabilities:
Anthropic's team, which had never done it before, stripped the model's refusal behavior for about $4,400 of compute with no meaningful loss of capability.
Cyber offense is now limited by energy, chips, and open-source model releases rather than headcount. Nation-states can run thousands of autonomous campaigns against smaller defense suppliers instead of targeting only the top fifty companies, while criminal groups can rent the same capabilities by the hour.
There are two proof points from the last three weeks:
The people vector: TA419. On October 1, Proofpoint detailed TA419, a China-aligned espionage group targeting U.S. and Japanese organizations since April 2025.
The systems vector: DIVD. On September 21, the Dutch Institute for Vulnerability Disclosure was breached by an autonomous AI agent, according to its incident response team. The agent exploited two previously unknown Zammad helpdesk vulnerabilities, gaining remote code execution and escalating to root access within seconds.
DIVD described the attack as loud and messy, with password spraying disrupting the agent’s own man-in-the-middle attack and code comments exposing its steps. Still, it gained root access in seconds, exfiltrated volunteer emails, and reached sensitive material before containment. DIVD’s monitoring, network segmentation, and incident-response team stopped the intrusion, but smaller suppliers lacking these defenses might never detect it.
The pause is not a reason to stop. It is a chance to prioritize. The recommendation is simple: achieve the full compliance baseline you are already obligated to meet under NIST 800-171 and DFARS 7012, then make sure the daily operational controls that contain a breach are in place and working.
Close the avenues of attack. DIVD highlights the need to inventory internet-facing assets, patch exposed systems within days, and retire those that can't be patched. TA419 reveals the human risk of convincing, malware-free phishing that steals sessions and bypasses traditional MFA. Defenses include employee training and phishing-resistant authentication, such as passkeys or hardware keys, for anyone handling CUI.
Assume something gets through and be ready. DIVD's story ended with limited damage for one reason: it detected the intrusion within a day, had segmented its network, and had an incident-response team that moved. Its response highlights capabilities organizations need to detect and contain attacks before they escalate:
None of this requires a C3PAO. All of it requires the CEO to decide it matters and fund it. If the November date was the only thing holding that decision in place, the adversaries described above are counting on you to let it go.
The next CMMC framework remains uncertain, with the task force’s report still unpublished. But the direction is clear: prioritize practical cybersecurity and scalable defenses over certification burdens, while easing access for smaller contractors.
An investment in vulnerability management, hardened identity, continuous monitoring, hunting, and response is not a bet on one version of CMMC. It does three things at once:
The parts of CMMC most likely to change are the assessment mechanics. The parts least likely to change are the ones tied to incident avoidance and business damage. Invest there.
The department made a defensible call to lower the compliance burden on the base. The adversaries did not make a matching call to lower the threat. They are automating, scaling, and coming after more of us in parallel than they ever could before. The pause gives you the room to spend on what actually stops them, but it does not give you permission to stop.
CMMC requirements may be changing, but the need for effective cybersecurity isn't. RADICL helps defense contractors strengthen their security posture through continuous threat monitoring, proactive defense, and hands-on compliance guidance.
Connect with our team to make sure your organization is prepared for today's threats and whatever comes next.