The 2026 Breach Crisis Demands Defense at Machine Speed

The 2026 Breach Crisis Demands Defense at Machine Speed

Governed AI can accelerate intervention while human judgment sets the boundaries and security providers prove their defenses work.

Chris Jacob, Field CISO, Securonix

 

By September 30, the Identity Theft Resource Center had tracked 3,064 publicly reported data compromises in 2026. The year continued on pace to set a new record.

IBM’s 2026 Cost of a Data Breach Report puts the average breach cost at $4.99 million, up 12%. Most of that bill came from detection and escalation, together with lost business: investigating what happened, determining the damage, and absorbing losses from disrupted operations and departing customers. AI-enabled malicious breaches were more expensive still, averaging about $6 million in the study.

I’m expecting breaches to keep escalating as attackers use AI to extend their reach and accelerate parts of their work. Stolen credentials and unpatched software will continue to give them the openings they need. AI can now add capacity to exploit those weaknesses, even as organizations introduce AI applications with access to sensitive information.

Governed AI can help teams investigate and interrupt threats at machine speed while human intelligence supplies the business context, skepticism, and judgment that determine whether an action serves the organization. Security providers have to make that relationship work under pressure, with controls customers can inspect and evidence that the defense actually intervened.

Waiting for an analyst to assemble every record and approve every routine step leaves valuable time available to an attacker and giving AI broad authority creates new exposure of its own. Teams need to establish permitted actions before an incident, then test how quickly their systems can execute them and when a person must take over.

 

Why Breaches are Escalating

Attackers are exploiting flaws organizations have yet to fix, including weaknesses whose full danger has gone unrecognized. Verizon’s 2026 Data Breach Investigations Report, based on 2025 data, identified vulnerability exploitation as the leading breach entry point, accounting for 31% of breaches. Verizon also described AI accelerating exploitation of known vulnerabilities and shrinking the time available to defend.

The criminal market has already compressed the work between gaining access and using it. Mandiant’s M-Trends 2026, based on its 2025 investigations, put the median handoff between initial access and a secondary threat group at 22 seconds, compared with more than eight hours in 2022. While that measures a specific handoff, rather than the time to complete a breach, access specialists were increasingly preparing the next group’s malware or tunnels in advance.

A convincing call to the help desk opens the route into cloud systems. Mandiant documented a shift toward voice phishing and the theft of long-lived OAuth tokens and session cookies. With usable stolen access, an attacker may be able to export data the same way a legitimate integration does. The Canadian Centre for Cyber Security described those methods in its May warning. The activity may look routine. Investigators will have to establish whether activity the application permitted served a legitimate purpose.

AI can help attackers repeat and refine these methods, while businesses are giving AI applications access to sensitive systems of their own. I expect heavy pressure on security teams as old flaws remain unpatched and stolen access stays usable longer than it should. We should be careful about assigning AI a precise share of the increase, though. ITRC found that 77% of public breach notices in Q3 omitted an attack vector. We often lack even a basic account of how an incident began, which makes confident claims about AI’s involvement difficult to defend currently.

 

The Investigation Itself

IBM attributed 63% of the average breach cost to detection and escalation, together with lost business. Companies are paying to piece together what happened while disrupted services and departing customers add to the losses. Investigators need records from across the organization to establish what was accessed and help business leaders decide which operations can safely continue.

The $4.99 million average gives leaders a starting point for discussing exposure risk. An organization’s own costs will depend on its circumstances, and multiplying IBM’s average by the ITRC breach count would mix different populations and reporting periods into a misleading national estimate. With the rapidly evolving threat landscape and how cybersecurity teams operate, we need to ensure our takeaways are grounded in reality.

IBM also found that identifying and containing a breach took an average of 247 days, up from 241 in the previous edition. That clock runs through containment and measures something much broader than how quickly an analyst reviews an alert. A team can clear its queue faster while an attacker continues using access its monitoring has missed.

However, AI can help gather records and connect activity that would otherwise take an analyst time to reconstruct. Evidence needs to stay with the investigation when an application owner or identity specialist joins, so they can assess the proposed response without starting over. Teams will also need to agree in advance who can revoke access or interrupt a service. Otherwise, time saved investigating can be lost waiting for someone authorized to act.

 

A Known Weakness is an Open Door

The October PHP findings put a concrete example behind existing weaknesses. Researchers using AI demonstrated attack paths through longstanding flaws in PHP’s core engine, prompting the October 21 release of fixes for two critical vulnerabilities known as PHP2Shell. Under the right conditions, an attacker could execute code on a vulnerable server without logging in. The PHP Foundation traced the first signs of one problem to 2024. Work on a fix had repeatedly stalled over engineering tradeoffs, including compatibility risks, and limited developer resources.

One flaw involved PHP’s handling of exhausted memory: a request could leave internal data partially initialized, then cause that invalid data to be used during cleanup. On Apache with mod_php, researchers demonstrated remote code execution before the website’s application code even ran. Fixing the website would still leave the underlying runtime exposed. AI helped researchers demonstrate how an old weakness could become a working attack, and the findings rippled through the internet. Experienced engineers still had to resolve the design tradeoffs and deliver a patch that organizations could deploy safely.

The PeopleSoft attacks documented this year show a software exposure surviving an apparent mitigation. Oracle described CVE-2026-35273 as remotely exploitable without authentication, potentially allowing remote code execution. Mandiant’s June research documented remote management agents, internal discovery, and attempts to move between PeopleSoft systems.

By September, Mandiant and Google Threat Intelligence Group had identified renewed exploitation by UNC6240, associated with ShinyHunters. Some operators had blocked the vulnerable Environment Management Hub path with web application firewall rules while leaving the software unpatched. The attackers requested /%50SEMHUB/ instead of /PSEMHUB/, encoding the letter P as %50. The rule matching the literal path could miss the request; the application server instead decoded it and reached the vulnerable component.

A screenshot of a rule proves only that someone configured it. The control must be tested against the request as the application processes it, including transformations performed by proxies and routing components. An emergency restriction also needs someone responsible for its limits and for replacing it with a durable fix.

For the investigation, preserve the original and normalized request paths and correlate them with host activity. Probing alone cannot establish compromise. Unexpected executable content or a newly introduced remote management process adds evidence that the attacker moved forward. Patching closes the entry point, but earlier access, persistence, and credentials exposed to the compromised process still require investigation.

 

The Valid Access Paradox

Criminal operators reach the same destination without exploiting the target application code. In May, the Canadian Centre for Cyber Security described campaigns using voice phishing, help-desk manipulation, and stolen integration tokens to reach enterprise cloud services. Attackers removed information through legitimate APIs and export functions, often without deploying malware.

Instructure’s explanation of the Canvas incident described another route to elevated access: an attacker submitted malicious code in a support ticket, and a cross-site scripting vulnerability allowed an authorization token to be obtained when a customer service agent opened it. In both examples, investigators had to follow authority the application could accept. Here’s where we see a valid account serving the attacker.

A usable stolen bearer token can allow requests without the victim entering a password again, while refresh tokens can support further access under the platform’s controls. Password resets therefore require application-specific verification. For example, Microsoft notes that Entra ID cannot directly revoke a session token issued by an application, because the application controls that session.

The same hard truth belongs inside the enterprise. ITRC tracked 21 insider wrongdoing events in the first half of 2026, compared with three in all of 2025. Counts are small, but they reinforce the need to examine how legitimate access is used. The center’s Q3 breakdown included system and human error alongside cyberattacks. Security teams can miss misuse and mistakes inside the organization when all their attention goes to outside attackers.

Behavioral evidence gives analysts a place to start, such as an account suddenly accessing unfamiliar data or exporting far more than usual. The person using it may have taken on new responsibilities or be carrying out an approved task, but analysts need to check that explanation against the audit trail and business records, then confirm it with the team responsible for the work before concluding the activity was malicious.

 

AI Is Giving Attackers More Capacity and Creates Targets of Its Own

IBM reported that one in four malicious breaches in its study were AI-enabled, with costs averaging about $6 million. Its detailed comparison puts AI-enabled malicious breaches at $6.04 million and malicious breaches without AI at $5.03 million.  The association does not establish how much AI alone added to the bill, but it gives security leaders a costly group of incidents to examine.

As we’ve seen, AI can help an attacker produce a convincing impersonation or examine code for weaknesses, and automate parts of an operation. IBM identified deepfake impersonation and AI-enabled malware as major components of the reported AI-enabled activity. I’d assume wider access to those capabilities will continue to help sustained breach pressure. Defenders will have to strengthen verification while reducing the time between recognizing malicious activity and stopping what the attacker can do next.

The PeopleSoft request-path bypass establishes how a control was evaded, but the published analysis does not establish AI involvement.

AI applications are becoming targets and carry access risks of their own. An agent that reads external material and invokes business tools creates an authorization boundary between the content it receives and the actions it can take. Microsoft’s guidance identifies risks involving malicious instructions in retrieved content, excessive permissions, and abuse of delegated authority. It recommends authorization on individual tool actions and limits on agent privileges.

Before deploying an agent, check that its permissions and tools fit the work you’re asking it to do. Keep enough context in the audit trail to investigate actions outside that assignment, while limiting the sensitive information those records contain.

 

Governed AI Within The Attack Window

Machine speed has a new practical meaning in security operations. An attacker using a stolen credential can finish exporting data while an analyst is still waiting for another team to send the logs. AI needs to help the team act during that window, with enough evidence to justify restricting access and permission to carry out the restriction. The investigation should then verify that the action stopped further access.

Our threat research team discovered TASK#STOMP, a Windows backdoor that automatically collects business documents and watches for new or changed files to upload. It uses familiar Windows components, including PowerShell and scheduled tasks, with multiple persistence mechanisms that can keep it running after one component is removed. Investigators need to connect those behaviors to understand the operation and determine what containment requires. While they assemble that picture, automated collection can continue. The response needs to stop further transfers and address the persistence that could bring the backdoor back.

Suppose an integration starts exporting a dataset it hasn’t accessed before. AI could review the requests made with its token, identify the issuing application, and check whether the activity fits the integration’s approved purpose. A scheduled migration might explain the export, so the investigation needs access to that business context.

The organization could permit a temporary, reversible restriction when the evidence meets agreed policy conditions. Limits should be enforced outside the model, so it cannot expand its own authority. The responsible team needs the supporting records and confirmation that the application actually blocked further access. Actions that could interrupt a critical service should require human approval, consistent with Microsoft’s guidance on checking authorization for each action and limiting agent permissions.

Human intelligence shapes the policy and investigates ambiguity. Analysts also need the authority to challenge an AI conclusion and change a failing procedure.

Prevention will always still begin with removing exploitable weaknesses and limiting access. Governed AI extends work teams can perform as an attack develops; but it cannot compensate for every missing control.

 

Consequences Before Disclosures

At the University of Illinois, the Canvas outage moved Friday final exams to Mother’s Day, May 10. The university’s May 9 announcement acknowledged religious observances and family plans, while explaining the problems that extending finals would create for travel, jobs, housing, and some instructors’ contracts. Rutgers University-New Brunswick also postponed its Friday finals. Academic leaders had to choose disruptive options while the security investigation continued.

Capcom’s October disclosures reported 11,774,247 affected CAPCOM ID records containing email addresses and account metadata, with smaller groups affected differently. Its follow-up identified access to e-Capcom pages for 19 of 196 accounts whose passwords had been changed without authorization. Purchase and inquiry histories were among the additional information exposed while credit card information was excluded.

These are all pieces of a puzzle for an impersonator or attacker. A recent purchase or unresolved inquiry could make a fraudulent refund or support message more convincing. These are plausible uses of the disclosed information, rather than reported follow-on attacks against Capcom customers. As a CISO, I would ask support and fraud teams whether their verification procedures rely on any of the exposed information. A breach gets a second life when stolen information can still pass as proof and build a credible story.

An August CISA advisory described a Gunra incident in which attackers deleted backup and archived data at both the primary data center and the disaster recovery center. Copies in separate locations remain exposed if the same compromised access can destroy both. Recovery exercises should assume production credentials have been stolen, then test whether the backups remain protected and whether the team can restore a working business service.

 

Prove How Your Defenses Intervene

We are moving into an era where security providers need to show how their defenses handle AI-powered attacks across the systems customers already run. Customers will need to understand what the system is allowed to do and see evidence that its actions worked.

During an evaluation, I’d use scenarios customers could encounter across their environment, such as an attempt to exploit an exposed application or an account unexpectedly exporting sensitive data. The provider should demonstrate how its system gathers enough evidence to support a response and carries out actions within the customer’s policies. I’d then check whether access was actually blocked and whether analysts could review or reverse a mistaken restriction.

IBM found that organizations extensively using security AI and automation had average breach costs $1.93 million lower and breach lifecycles 65 days shorter than organizations using none. A buyer still needs evidence of what a particular product can deliver in their environment. I’d ask any provider to demonstrate an investigation through completion, including how an analyst can catch an incorrect recommendation and prevent it from disrupting the business.

 

Expectations of Security Leaders Forward

I expect CISOs to get much tougher about the AI they buy as breach pressure grows. Providers have to demonstrate what their systems can do and how permissions hold when the evidence is incomplete. Human oversight will need to be built into that workflow, with clear responsibility for decisions that could disrupt the business. Analysts will have to see the evidence behind a recommendation and be able to stop or override an action without reconstructing the investigation.

Security teams will look more closely at how people and AI agents use their access. An employee and an agent may be able to reach the same sensitive data, sometimes through service accounts or integrations. Teams need to check whether that activity fits the work each was authorized to do. When they restrict access, they also need to confirm that the application actually enforces the change.

I also see more pressure on providers to measure their contribution to preventing harm. A credible evaluation should account for errors and operational disruption alongside speed. Customers need to know how much work the defense completes and which decisions still depend on their own teams.

Other defenders may be dealing with the same exposure your team just investigated. Share the technical findings that helped you act, with sensitive details protected and gaps in the evidence clearly marked. Explain which finding changed the response so another team can use it to test its own defenses.

Before the next budget review, run an attack exercise against a critical service with your provider and the people authorized to protect it. Agree on what AI can do under policy and which decisions need human approval. Then check whether the defense actually interrupted the attack and left a usable record of its actions, with the service still running. Give every failure an owner and a deadline. Bring the unresolved failures to the people approving the spend, and make fixing them part of the commitment they are buying.

 

Sources

  1. WBAY reporting on the 2026 record breach pace
  2. ITRC third quarter 2026 data breach update
  3. Help Net Security reporting on IBM breach costs and AI attacks
  4. IBM July 29 announcement of its 2026 breach cost study
  5. IBM analysis of AI adversaries and 2026 breach economics
  6. Verizon 2026 DBIR announcement
  7. Mandiant M-Trends 2026
  8. Canadian Centre for Cyber Security alert AL26-010
  9. IBM Cost of a Data Breach Report 2026
  10. Oracle Security Alert for CVE-2026-35273
  11. Mandiant June PeopleSoft campaign analysis
  12. Mandiant September 25 PeopleSoft campaign analysis
  13. Instructure customer FAQ on root cause and investigation evidence
  14. MITRE ATT&CK application access token theft
  15. Microsoft guidance on revoking user access
  16. ITRC first half 2026 data compromise findings
  17. Microsoft AI agent shared responsibility model
  18. Microsoft guidance on indirect prompt injection defenses
  19. University of Illinois May 9 restoration and revised exam schedule
  20. Rutgers New Brunswick May 7 postponement of Friday finals
  21. Capcom October 22 incident disclosure
  22. Capcom October 23 incident update
  23. CISA August 10 Gunra ransomware advisory announcement
  24. PHP Foundation October 21 account of the PHP security release
  25. Wiz PHP2Shell research and AI assisted vulnerability analysis

 

 

See Securonix in Action

See how Securonix brings together AI-powered threat detection, intelligent automation,
and advanced analytics to help organizations identify threats faster and respond with confidence.

Let us show you what’s possible.