A threat report says MFA is being bypassed faster than insurers can reprice it, and AI agents are doing the scanning
By Albina Bieci, Staff Writer
Session hijacking, device-code phishing and token harvesting are getting past the control most cyber policies still treat as the baseline. Autonomous intrusion agents are reaching ransomware in two to three weeks.

NEW YORK
A quarterly threat report from Beazley Security, analyzed this week by Insurance Business America, warns that threat actors are systematically bypassing multi-factor authentication through active session hijacking, device-code phishing and infostealer token harvesting, while autonomous AI intrusion agents scan and compromise unpatched servers at machine speed and deploy ransomware two to three weeks after initial access.
The report's second finding is the one with commercial consequences: organizations that rely on legacy MFA checklists, without a tested incident response plan, carry operational gaps their cyber policies may not answer for.
What the report says
The report identifies three MFA bypass techniques in active use: session hijacking, device-code phishing and infostealer token harvesting. It describes autonomous AI-driven intrusion agents scanning and compromising unpatched servers at machine speed, with ransomware deployed two to three weeks after initial access. It warns that organizations relying solely on legacy MFA controls face severe uninsured operational gaps without tested incident response plans.
The report does not, in the summary available, quantify how many incidents used each technique, name specific victims, or state which industries are most affected. Those details will be added here if the full report is published.
Why MFA stopped being a wall
Multi-factor authentication was designed around a specific attack. Someone steals a password, tries to log in, and cannot, because the second factor is a code on a device they do not have. That defense works, and it is why MFA reduced account compromise so sharply for a decade.
The techniques in the report do not attack the second factor. They avoid it.
Session hijacking works because authentication happens once and then produces a token, and the token is what every subsequent request carries. If an attacker can take the token, they do not need the password or the code. They are inside an already-authenticated session, which is why this technique is also called adversary-in-the-middle: the user authenticates to a proxy that passes the credentials through to the real service and then keeps the resulting session.
Device-code phishing exploits a legitimate feature. The OAuth device authorization grant exists so a device without a keyboard, a television or a printer, can be authorized by visiting a page on a phone and typing a short code. Attackers send the code to a target and ask them to enter it, and the target does exactly what the flow intends: they authenticate to the real provider, satisfy whatever MFA the provider requires, and approve the attacker's device. No fake login page, no certificate warning, nothing to notice.
Infostealer token harvesting is the third path and the least exotic. Malware that infects a workstation takes the browser's stored session cookies along with saved passwords, and the attacker replays those cookies from their own machine.
The common thread is that all three end with the attacker holding a valid session that was legitimately created. The MFA prompt fired. The user approved it. The control worked as designed and the account was still compromised.
What actually closes those paths
The controls that address these techniques are not replacements for MFA. They are the conditions that make an MFA prompt mean something.
Phishing-resistant authenticators, meaning passkeys and hardware security keys bound to the origin, are the direct fix, because they cannot be relayed through a proxy. Conditional access that binds a session to a device posture and a location, so a token replayed from an unfamiliar host fails. Short session lifetimes with re-authentication for sensitive actions. A prohibition on device-code flows unless they are explicitly required, and monitoring for their use. And detection that looks for the anomalies rather than the logins: impossible travel on a session, a mailbox rule created minutes after a sign-in, a mass download from a file share.
None of that is exotic. Most of it is configuration in tools organizations already own, which is why the report's framing is about checklists rather than technology.
Machine-speed scanning and the two-week dwell time
The second finding changes the economics of patching.
Human-driven intrusion has a rhythm. A scanner runs overnight, a person reviews results the next day, a decision is made, an exploit is attempted. That rhythm created a window of weeks in which an unpatched system was exposed but not yet exploited.
An autonomous agent collapses that window. It scans, identifies a vulnerable service, selects an exploit, attempts it, and moves to the next host without waiting for anyone. When the same automation is pointed at a known vulnerability with a public exploit, the population of exposed systems is being worked through in hours rather than weeks.
Two to three weeks from initial access to ransomware is consistent with what incident responders have been reporting: the dwell time that used to be measured in months is now measured in days, and the reconnaissance, lateral movement and staging phases have been compressed into a single automated run.
The defensive implication is not that patching matters more. It is that the interval between a patch being available and it being applied is now the exposure window, and an organization that patches on a monthly cycle is running a month-long exposure on every critical vulnerability that gets a public exploit.
Why the insurance gap is a documentation problem
Cyber policies have increasingly attached conditions to coverage: minimum security controls warranted at inception, MFA required for remote access, backups maintained offline, an incident response plan in place. Some of these are warranties, meaning a breach of the condition can affect the claim itself rather than merely the price.
The report's warning about uninsured operational gaps is pointing at a specific mismatch. An insured organization can hold a policy that warrants MFA, satisfy that warranty literally, and still be breached by an attacker who never broke the MFA. The coverage question then turns on what the policy says about the technique rather than about the control.
The practical answer is a review, done before a claim rather than after one: read the warranties, confirm the controls they name are actually deployed as described, document that they are, and check what the policy requires in the first hours of an incident, because notification timing and approved vendor requirements are the conditions most often missed.
What a tested plan looks like
A tested plan is not a document. It is a set of decisions that have already been made, by people who know who makes them.
Who declares an incident, and who can make that call at three in the morning. Who calls the insurer, on what number, within what window. Which forensics firm is on retainer and how quickly they can be engaged. Whether the organization will pay a ransom, decided in advance by the people with authority to decide it. How communications with customers, regulators and staff are handled, and who speaks. Where the backups are, whether they have been restored in a test, and how long a full restore takes.
The last item is the one that separates organizations that recover from those that do not, and it is the cheapest to verify.
The physical side of the same report
For organizations that run both security officers and networks, the report describes a convergence rather than a separate problem. Access control, camera systems, alarm panels and building management systems sit on the same networks as everything else, and they are frequently the least patched devices in the estate because nobody owns them.
An intrusion agent that reaches a camera server has reached a physical security system, and the response requires someone who understands both.
Five questions a board should be able to answer this quarter
The report is written for risk professionals, and it translates into a short list of questions that a board or an ownership group can put to management without any technical background.
Which of our authentication methods can be relayed by a proxy, and what percentage of our workforce still uses one. Are device-code flows disabled for every application that does not require them. How long is a session valid, and what would an attacker holding a stolen token be able to reach before it expires. What is our median time from patch release to production for a critical vulnerability with a public exploit. And when did we last restore our backups in a test, rather than verifying that a backup job reported success.
Each answer is a number or a date. An organization that cannot produce them is the organization the report is describing.
Related coverage
The Watchline has reported on the ransomware economy this month. An alleged INC and Lynx affiliate was extradited from Italy and appeared in a Tennessee court. Officials also ruled out AI and cyberattacks in a week of FAA, United and NSA outages, a reminder that attribution takes time even when the disruption is immediate.
The service connection
USPA Nationwide Security provides corporate risk intelligence, cybersecurity assessment and converged physical and cyber security review, including incident response planning support and vendor exposure analysis. Requests can be made through the corporate risk intelligence page, the cybersecurity page, or by calling (800) 214-1448.
Sources and status
Reporting on the report is attributed to Insurance Business America. The full methodology, incident counts and industry breakdown remain open pending publication of the underlying Beazley Security report. Technical references include CISA and the NIST Cybersecurity Framework.
Services in New York, NY
Sources
About the reporter
Staff Writer
Staff writer for the USPA Nationwide Security newsroom. Receives newsroom tips, corrections and editorial inquiries.
Author biography and credentialsMore from The Watchline

Industry
Chase, Visa and Mastercard traced their exposures to vendors, and the White House named a cyber director the same day
Chase pointed to marketing data, Mastercard to a cardholder benefits program, Visa to an Asia-Pacific vendor with no card numbers involved. All three investigations are running against an extortion claim.
· Albina Bieci

Industry
Officials ruled out AI and cyberattacks in a week of FAA, United and NSA outages, and Congress wants the evidence
Three unrelated disruptions, one week, and a public that cannot tell a bad deployment from an intrusion. Officials ruled out AI and cyberattacks; what they have not yet published is the root cause of any of it.
· Albina Bieci

Industry
USPA expands its cybersecurity practice to website security upgrades and secure rebuilds
The practice now scopes hardening work on existing sites alongside full secure rebuilds, hosting and email migrations, and continuing management for businesses whose websites have outgrown the platform they were built on.
· Albina Bieci





