A credo for the AI era: Trust, but Decompile

COMMENTARY: For 20 years, enterprise security has run on a simple bet: Find the vulnerability. Score its severity. Patch on schedule. Repeat.

We built whole organizations around that ritual. Vulnerability management teams, change advisory boards, maintenance windows blocked out a quarter in advance.

[SC Media Perspectives columns are written by a trusted community of SC Media cybersecurity subject matter experts. Read more Perspectives here.]

And yet the breaches keep coming. Verizon's 2026 Data Breach Investigations Report puts exploitation of software vulnerabilities at the top of the list of ways attackers get in, up 31% year-over-year, while credential-based attacks actually declined. After two decades, even though we've organized ourselves to prevent exploits, they keep happening.

The disconnect isn't subtle. Modern security teams are still organized around cataloging and scoring software flaws. What's the CVE? What's the CVSS? Was it added to the KEV? If so, how wide is the patch window? Attackers are answering a different question entirely, and it's a better one: What’s actually running in this environment, right now? Can I reach it?

Breaches live in the space between those two questions.

That space keeps getting bigger, and AI scales it. J.P. Morgan's Michael Cembalest laid out the arithmetic in his July 2026 Eye on the Market note, drawing on the bank's Patchmageddon research: Roughly 48,000 vulnerabilities disclosed in 2025. Roughly 7,500 patched.

The median gap between disclosure and first exploitation, which ran close to a year as recently as 2021, has collapsed to about one day, and Cembalest expects it to reach a single minute by the end of 2027. A model can read our vendor's patch, work backward to the flaw it fixes, and produce a functioning exploit in less time than it takes our change advisory board to find a room. Our remediation has become the attacker's spec sheet. And there’s no human anywhere in the attacker’s loop. 

Once exploitation reliably beats disclosure – a threat that arrives before the advisory, let alone the patch– we cannot schedule our way to safety. We should no longer ask: How bad is the flaw on a scale of 1 to 10?

The question has become narrower and considerably less comfortable: Is this specific vulnerability exploitable, in our environment, given what’s actually deployed, right now?”

Our industry has not been able to answer that question. So we built three stand-ins and agreed not to look at them too closely.

First, we have the scan-and-patch program, which relies on public catalogs like the National Vulnerability Database (NVD). Match our components against the NVD and similar feeds, then work the list. This approach held up reasonably well when the catalog kept pace with the code. It no longer does.

Research into flaws discovered by AI, including Anthropic’s Mythos, found that the overwhelming majority carried no public advisory at all. That means scanners that only recognize what's been publicly catalogued will report “all clear” while staring directly at a code flaw that's going to get us.

Second up was provenance: trust the source, not the artifact. Sign our code, verify the signature, and sleep well. The trouble here: a signature tells us who vouched for something, not what it does. The SolarWinds Orion update was signed. Its provenance checked out. It was also backdoored, and it walked into thousands of environments on the strength of that signature.

We could call SolarWinds an outlier, and at that scale it was. Then look at open source, where it isn't an outlier at all. Sonatype counted more than 450,000 malicious open-source packages in 2025, a great many of them published through entirely legitimate maintainer accounts that had been quietly taken over. The trusted channel has become the preferred channel, for the obvious reason.

Third on the list has been source analysis — SAST, composition tools, anything that reads our code and manifests. These are genuinely useful to developers and genuinely blind as a deployment control, because source represents the blueprint and not the building. Endor Labs found that roughly 95% of open-source vulnerabilities live in transitive dependencies that most manifests never surface. 

The common failure here with all these stand-ins: they are descriptions of software: A catalog entry, a signature, a manifest. None of them is the actual software. The attackers stopped trusting descriptions of software a long time ago. They read the binary. We should, too.

What does that mean? Open the software artifact — the actual container, package, or binary that will execute in the company’s environment — and find out what's inside it.

Our team have taken to updating the Reagan-era credo: “trust but verify” – for our modern, software-powered era. The new slogan: “Trust, but Decompile.” In other words: Take the code apart the way an attacker would and learn what it does when it runs. That’s a different subject from what its paperwork claims about it.

This only works as a loop with three stations:

None of these steps requires ripping anything out because nobody will. But, by gating our CI/CD pipeline and artifact registry with binary verification on whatever will ship, we can attest to the security of every release that passes. 

When it comes to deployment, harden the company’s most exposed systems first — internet-facing assets, regulated data, and the handful of workloads that would make the news. Then re-verify and re-attest the hardened result, wiring the attested baseline into the NDR and EDR telemetry we already pay to collect.

Doing that means the next disclosure gets triaged on whether the bug was present and reachable in our environment, rather than on a score somebody with no knowledge of our environment assigned it in a lab. With such a system, the flood of code drift and new vulnerability disclosures sorted by severity gets reduced to a short list of tickets with relevance to our environment and evidence attached.

This does not make it so we can prove we have safe software. Nobody can certify the absence of a flaw nobody has thought to look for yet. Please show anyone selling that certainty the door.

In reality, we’ve just retired the patch calendar in favor of a continuous loop that runs on the adversary's clock instead of ours. And, these days, when we keep up with our adversaries, we stay safe.

Mario Vuksan, chief executive officer, ReversingLabs  

SC Media Perspectives columns are written by a trusted community of SC Media cybersecurity subject matter experts. Each contribution has a goal of bringing a unique voice to important cybersecurity topics. Content strives to be of the highest quality, objective and non-commercial.