FBI Certifies Death of Security Through Obscurity: AI Finds Bugs Where None Were Expected

On September 13, 2026, Brett Leatherman, assistant director of the FBI Cyber Division, made public a finding researchers had feared for months: AI models have uncovered significant vulnerabilities in open-source libraries considered secure after ten years of stress-testing. The confirmation arrives as Microsoft closes a record Patch Tuesday with 974 CVEs, including components untouched for years: Telnet client, Windows RNDIS, 1980s-era NFSPortmapper, Link Layer Topology Discovery inherited from Vista.

The news isn't that AI finds more bugs. It's that AI finds bugs in places no one thought to look, making economically viable what was previously prohibitive due to a scarcity of expertise. The defensive paradigm built on the assumption that certain systems remain unknown to attackers is clinically dead.

AI Models Have "Broken" Decade-Old Open-Source Libraries

Leatherman's statement is precise and without appeal. According to The Register, the FBI verified that "the latest models were able to break those and say, 'yeah, there's significant vulnerabilities in here'" referring to open-source libraries powering roughly 80% of global web servers. Ten years of maturation is no longer a sufficient guarantee.

The mechanism is simple to describe and difficult to counter. AI agents can parallelize reconnaissance and investigation at a scale impossible for human teams, driving the marginal cost of analyzing obscure systems, proprietary protocols, or legacy code toward zero. What protected these systems wasn't architectural robustness, but information asymmetry: familiarity with specific tech stacks was confined to a handful of people. As John Hultquist of Google Threat Intelligence Group put it, "they've been largely secured because the expertise was in a handful of people's heads, and that's not going to last forever."

The convergence of sources is clear. Dustin Childs of Trend Micro's Zero Day Initiative observed that in the September 2026 Patch Tuesday, "when you look at all of the components patched by Adobe and Microsoft yesterday, you see components no one has talked about in years." The list is archaeological: Telnet, a 1969 terminal emulation protocol; RNDIS, Remote NDIS for USB; NFSPortmapper, an 1980s Unix port-mapping service; LLTD, a topology discovery protocol introduced with Windows Vista. None of these components are technically new. All are back in the crosshairs because AI made it worthwhile to look.

Patch Gap Compressed to Hours and the Offensive-Defensive Asymmetry

Temporal compression of exploitation is the second pillar of the imbalance. At least four espionage crews — most with suspected links to China — have reduced the Chromium patch gap to hours. The mechanism is surgical: develop exploit kits in the window between upstream patch release and downstream stable distribution, when systems are technically fixed but operationally still exposed.

Critical infrastructure vulnerability is equally documented. Five US agencies reported in a joint advisory that attackers used AI-generated exploitation scripts to breach internet-exposed Siemens S7 Series PLCs. The joint statement is categorical: "this is not a theoretical risk – it is an active threat." Hultquist emphasized AI is particularly dangerous in ICS/OT because it excels at technical troubleshooting and knowledge of systems combining obsolete technologies with modern exposure.

The problem is structurally asymmetric. The attacker faces no governance constraints, approval processes, or change management. As Venture in Security notes, AI "is completely changing the economics of attacks by reducing the marginal cost of reconnaissance, investigation, and attack preparation." The defender must maintain compliance, test, and deploy. The speed of vulnerability generation outpaces remediation, and data on AI-generated patching quality offers no shortcuts.

"Security architectures built around the assumption that attackers won't notice something will become increasingly fragile" — Venture in Security Analysis

The Failure of AI-Generated Patches: The 1Password and Veracode Numbers

Attempts to close the defensive gap with the same offensive tools show disappointing results. According to a 1Password study cited by The Register, 6,080 patches were generated with ChatGPT-5.5 and Opus 4.8 for 6 CVEs. The full success rate — patches that resolve the vulnerability without altering application behavior — was 26.0%. The failure rate, defined as patches that do not resolve or introduce new vulnerabilities, reached 53.9%.

A parallel Veracode study, cited in the same source, records an average security pass rate of 56% for AI-generated code, calculated across over 100 models and 80 coding tasks. The dual data points converge: automatic generation of secure software is significantly harder than automatic generation of exploits. The asymmetry is not temporary nor solvable with more powerful models in the near term; the two studies use already advanced architectures (ChatGPT-5.5, Opus 4.8) that may be surpassed, but without evidence that successors will reverse the trend.

Katie Moussouris of Luta Security summarized the situation with diagnostic precision: "if there is something to find, they will find it." The phrase is not a prediction. It is a statement of the annulment of the resource constraint that served as a natural limit on offensive investigation.

What to Do Now

• Reduce surface exposure: evaluate elimination or isolation of legacy components no longer actively maintained, rather than assuming they remain unknown.
• Recalibrate security by design: shift resources from reactive vulnerability management toward architectures that do not depend on obscurity for their robustness, consistent with frameworks proposed by industry analysis.
• Reconsider the OT/ICS perimeter: treat internet exposure of industrial systems as a calculated risk in light of AI's capacity to generate exploits specific to PLCs and proprietary protocols.
• Invest in enterprise resilience: prioritize response and recovery capabilities over the patching race, given evidence that defensive remediation speed does not scale with offensive generation speed.

Meta-Hacking and the Compression of Attack Time

A specific case illustrates the depth of the imbalance. Varonis Threat Labs discovered CoSnitch (CVE-2026-24301, CVSS 8.8 HIGH) using a technique the industry calls "meta-hacking": the AI was induced to reason about its own architecture, revealing without reverse-engineering a ?autorun=1 parameter that exposed the system. It is the third Copilot flaw discovered by Varonis in 2026, all sharing a "one-click" pattern.

The CoSnitch case is not anecdotal. It confirms AI can be used to extract vulnerabilities from the model itself, further shortening the discovery cycle. The dossier does not specify whether this technique is systematically replicable on other AI assistants, but the principle is established: obscurity does not protect even the AI systems generating the offense.

Cloud Security Alliance CISO summits — San Francisco in May, New York and Washington in June 2026 — identified the trend with clarity. According to testimony gathered under Chatham House Rule, AI-enabled vulnerability discovery "will break existing governance and remediation processes." The statement is qualitative, not quantified, but the convergence of enterprise CISOs on this point is notable: the vulnerability management system as structured today cannot withstand the volume and velocity of the new offensive economy.

Why It Matters

Security through obscurity was never a robust strategy, but it was a functional one for decades. AI has annulled its economic premises. The finding does not belong to columnists or bloggers: it comes from the FBI Cyber Division, Google Threat Intelligence, Trend Micro ZDI, and a joint advisory from five federal agencies. The paradigm is dead for arithmetic reasons, not ideological ones.

Time is tightening in opposite directions. For the attacker, the cost of discovering vulnerabilities in obscure systems tends toward zero. For the defender, the cost of patching them remains high, and attempts at defensive automation fail more than half the time. The gap is the maneuvering space for those who must decide where to invest. The clinical reading suggests the answer is not more tools, but architecture: less surface, fewer assumptions, less trust in obscurity as a defensive property.

Sources

• https://www.theregister.com/security/2026/09/13/security-through-obscurity-is-dead-and-ai-delivered-the-fatal-blow/5296000
• https://ventureinsecurity.net/p/how-ai-is-actually-remaking-security
• https://cloudsecurityalliance.org/artifacts/ai-security-through-the-ciso-lens
• https://ventureinsecurity.net/p/sponsor
• https://hubs.ly/Q04tPt5G0
• https://www.amazon.com/Cyber-Builders-Essential-Building-Cybersecurity/dp/173823410X/

Information verified against cited sources and current as of publication.

Sources