By Padraic O’Reilly, Founder & Chief Innovation Officer, CyberSaint
For twenty years, security teams have gamed out attack likelihoods using measures of sophistication. State sponsored actors could produce multi-stage campaigns with long chains of ttps and script kiddies couldn’t. That assumption is why a risk register ranks a multi-stage intrusion below a phishing email, and why a plant manager can defer a patch for years and call it accepted risk. In 2026, AI agents took it apart. The barrier to entry is now flimsy. Here is what I expect in 2027 for the people who run and defend infrastructure.
1. Likelihood scores will finally move, and underwriters will be the ones moving them.
In September 2026, researchers documented a single operator who pointed open-source agent frameworks at the web and breached at least 27 companies for about $25 of model spend per target. Most risk models still rate a chain like that as unlikely due to complexity. In 2027, I expect cyber insurers to ask a question most risk programs can’t yet answer: did likelihoods change when the cost of attacking fell, and if not, why not? Organizations that can show which exposures are actually reachable, and what each would theoretically cost to exploit, will get better terms than those that can only produce a heat map.
2. AI-executed intrusions’ rate will tick up, and 2027 will pass 2026’s total by midyear.
Our analysis of the AI Incident Database counts at least a dozen separate 2026 reports of AI agents under evaluation or training reaching real third-party systems. Maybe more importantly, the broader count of AI systems carrying out intrusions themselves has climbed from one in 2024 to four in 2025 to more than 30 in 2026. One recent entry: a China-linked espionage group used an AI penetration-testing framework against more than 200 organizations, with nine confirmed breaches. Agents excel at tirelessly chaining weaknesses that each look minor into sum-of-parts path that is much more significant. My falsifiable call: the 2027 count passes the 2026 total before the end of June. The defensive answer is to manage vulnerabilities as paths, not the old standby lists: which assets an attacker can reach, which single change breaks the most routes, and what the impact looks like across multiple chains.
3. The agent’s toolbelt will become the new software supply chain, and it will be poisoned.
Agents now acquire skills, plugins, and tool connectors the way browsers acquire extensions, often unreviewed and holding credentials that can read source code, touch production, or push a release. In 2025, attackers published malicious npm packages built to hijack locally installed coding agents. In 2027, I expect a widely installed agent skill or connector to become a multi-company breach, one that package-level SBOMs won’t see. The important inventory is what each agent is allowed to install, which connectors it trusts, and whose identity it runs under.
4. Agent identity will be the year’s big access-control problem.
More agents now do their work by driving a human’s screen inside a session that has already passed MFA and device checks. Every control in the stack sees a verified user on a verified device. I expect 2027 to bring a significant breach through exactly that gap, and auditors to respond by asking for something most organizations don’t have: a register of every agent, the identity each runs as, and who can switch it off. The fix: give each agent its own identity, limit its privileges, not just where it can go, and treat “the agent borrowed my session” as an access decision.
5. “We can’t patch that” will need evidence behind it, in the datacenter as well as on the plant floor.
Every infrastructure team has systems it has quietly agreed not to patch: legacy hypervisor hosts, end-of-life appliances, controllers nobody wants shut off. That was an accepted practice when exploitation trailed disclosure by weeks. Mandiant now estimates that, on average, exploitation starts about a week before a fix is available. In June 2026, CISA replaced fixed federal patch deadlines with a test based on internet exposure and whether exploitation can be automated. In 2027, I expect auditors and insurers to carry that logic into the private sector: every deferral will need proof that the asset can’t be reached, or that a control breaks the path to it.
6. Post-quantum will be a procurement and diligence issue long before it is an attack.
No one will break production encryption with a quantum computer in 2027. But from January 1, 2027, NSA requires new national security system purchases to support its post-quantum algorithms, and that rule will flow through the defense supply chain and then into commercial procurement. I expect post-quantum questions in acquisition due diligence and vendor questionnaires by the end of the year. The sensible first step is an inventory of where your cryptography lives and which data must stay secret longest.
What links all six is that the old shortcuts for estimating risk, from “too sophisticated to be likely” to “too awkward to patch” to “the user is verified”, are failing at once. At CyberSaint we’re betting that the next generation of risk management is continuous and grounded in what an attacker can actually reach, not what a checklist says, and I expect the industry to follow. The organizations that come out ahead in 2027 won’t be the ones that predicted every attack. They’ll be the ones that stopped treating sophistication as a proxy for likelihood.
##
ABOUT THE AUTHOR
Padraic O’Reilly is Founder and Chief Innovation Officer of CyberSaint, where he leads research on AI-driven cyber risk quantification, attack-path reachability and the governance of agentic AI.






