By David Lindner, CISO and DPO, Contrast Security
In October, researchers using AI found critical flaws in PHP that had been in the codebase for more than ten years. PHP runs about seventy percent of the web. The same week, Node.js, Python, and Microsoft all rushed out patches for an out-of-bounds write in Google’s Brotli compression library, a bug that had been fixed in the source code back in February without anyone recognizing it as a security issue. And a certificate authority trusted by every major browser was caught using an encryption key that was the literal number four. Nobody who has worked in this field for long was surprised by any of it. We have known for twenty years that the software underneath everything is held together with duct tape, popsicle sticks, playdoh, plausible deniability, and hope. What changed is that an attacker no longer needs talent or patience to find the tape.
Finding vulnerabilities was never the hard part. We built an entire industry around it. Scanners, SAST, DAST, SCA, IAST, pentests, bug bounties, and now AI agents that do all of the above while you sleep. Every one of those tools found more and in most cases they found different things. Every year the backlog got bigger. Andreessen Horowitz surveyed one hundred CISOs at companies over a thousand employees and found eighty seven percent expect budgets to rise in 2027, with most of the money going to tools that discover more. One fintech CISO in that survey said the quiet part out loud. AI finds vulnerabilities faster. It does not patch them faster.
Boards already ask how many vulnerabilities you closed and some may even ask what your mean time to remediate (MTTR) was. They have asked this for years as a standard practice, but I would guess most boards saw this is risk management vs vulnerability management. During this time, nothing has really improved, because the discipline built to answer it is broken at the root. Vulnerability management treats every finding as a unit of work to burn down. The chart goes up and to the right forever, and we call that a program as long as we can show the burn down. In 2027 that model collapses, security teams will force it. When AI can find and chain flaws in your custom code at machine speed, the list is infinite by definition. You cannot burn down infinity. The burn-down chart goes the way of the Gantt chart, something you might only show auditors and nobody else.
What replaces it is risk management, and I mean that as a focus and discipline. CVSS will be the first to go. A worst case scenario, subjective, severity score assigned by someone who has never seen your architecture tells you nothing about whether a flaw is reachable, exposed, or being probed right now. Bain estimates AI compresses a typical attack sequence from about four weeks to about eighteen hours. Eighteen hours is not enough time to triage a spreadsheet of nine 9.8 CVSS scores, which is what leads teams to just fix now in today’s world. The only input that matters at that speed is runtime context, meaning which code is actually executing in production, which paths an attacker can actually reach, and which ones are actually being hit. Fixing is only possible when that list is short and prioritized, and the list is only short when it is real. Companies that already instrument their applications in production will have that list. Everyone else will be patching as they do today through arbitrary “risk” scores.
Risk management also changes what gets protected. Every enterprise is running software it has forgotten about. An internal portal from 2014 running an end of life version of Java. A vendor integration that was supposed to be temporary. A microservice whose owner left in a reorg. Attackers ignored these because finding them cost money and their human time. AI makes finding them cheap with no real human time needed. The most expensive breach of 2027 will start in an application nobody on the security team knew existed, and the chain will have three low-severity flaws no scanner would have escalated. A vulnerability program would have rated each one a medium and moved on. A risk program asks whether that forgotten portal can reach the customer database, and if it can, shuts it off.
Agents force the same shift on the identity side. The a16z survey found fifty five percent of enterprises already run MCP servers or similar agent-to-application integrations in production. Forty six percent list agents exceeding their permissions or being compromised among their biggest fears. Those two numbers describe the same companies. We already watched OpenAI’s agents break out of a cybersecurity evaluation and work their way into another company’s production systems. That was a lab. In 2027 it happens at an enterprise, either through a prompt injection buried in a document or a web page that the agent was asked to summarize or because someone handed it production credentials to close a ticket faster. EDR, DLP, and Zero Trust all assumed a human was at the keyboard. An agent acting with your permissions looks exactly like you to every one of those tools. Treating agents as identities with their own blast radius is a risk decision. Treating them as features of the user who launched them is how the breach happens.
The doomsday scenario most people imagine is a grid going dark or a bank going down. I think the real version is quieter. Thousands of small breaches at thousands of companies, each traced to a flaw that had been findable for years and finally got found, none of them big enough for a headline. Insurers will notice before the press does which will then hit during policy renewal for everyone.
So in 2027, retire the traditional vulnerability management program and the burn-down chart with it. Finding and fixing does not go away. It gets absorbed into the development workflow, where AI harnesses scan, patch, and verify as part of the build the same way tests do today, and nobody calls that a security program any more than they call CI a quality program. The security control point moves later, into production runtime, where you can see what is running, what is reachable, and what is under attack, and prioritize on that instead of on a severity score. Replace the CVSS filter with reachability and attack data from production. Inventory every application you run, including the ones you wish you didn’t, and decide which ones can touch something that matters. Give every agent in your environment its own identity and its own permissions, and log what it does separately from the human who started it. Fund the engineering time to fix at the same rate you fund the tools to find, or stop pretending the discovery budget is a security investment. The attackers already made the switch from finding to exploiting at scale. 2027 is the year defenders either make the switch from counting to deciding, or keep counting while the count stops mattering.
