Keep the lessons.
Revisit the predictions.

THE YEAR IN CASE FILES

Trust was an operational dependency.

On 30 December 2025, the US Department of Justice announced guilty pleas by two cybersecurity professionals for participation in ALPHV/BlackCat ransomware attacks. Their professional experience makes the case especially troubling. The announcement does not establish that they were the incident responders hired by each victim they attacked.

The governance lesson is to control consequential access even when a person has relevant expertise and an urgent job to do. Incident response arrangements should establish authorisation, logging, expiry, and conflicts of interest before an emergency. Trust in professional competence and control over access address different questions.

Another form of delegated authority became visible in Cloudflare’s account of the Salesloft Drift incident. Compromised integration credentials were used to access Salesforce information. The relevant boundary was a third-party connection and its permissions; describing it simply as a failure to choose a strong user password would miss the mechanism.

A business should know what an integration can read or change, who owns it, and how it can be revoked without waiting for its original installer. The approval made when an application was connected may still be operating long after the original business need has changed.

HELPFUL INSTRUCTIONS / HARMFUL AUTHORITY

Support processes can carry an attacker’s request.

ClickFix-style attacks ask a person to carry out a harmful action under the pretext of solving a technical problem. The wider insight is behavioural: a convincing support instruction can borrow authority from the user’s desire to repair an error and return to work.

Help-desk impersonation poses a related problem. If a caller can persuade support to reset a consequential account, the password itself is only one part of the control. The recovery process, identity evidence, approval route, and ability to refuse an exceptional request matter as much as ordinary sign-in.

Training still has a role. It should be connected to an easy way to verify support instructions, a policy that permits a pause, and technical restrictions on consequential actions. Calling all awareness training obsolete would discard a useful control while leaving the approval process unchanged.

AI / CONTENT & INSTRUCTIONS

An assistant can cross a boundary the user cannot see.

The EchoLeak case study describes CVE-2025-32711, a zero-click prompt-injection vulnerability in Microsoft 365 Copilot. Researchers demonstrated how malicious content could cause information to leave the assistant’s intended boundary. The public research concerns a flaw in a production product; it is not evidence that a named customer suffered the demonstrated theft.

The important distinction is between information the assistant should examine and authority it should obey. Content from an email or retrieved document should not acquire the power to redirect an assistant merely because the assistant can read it.

For deployment, ask what the assistant can access, what it can send or change, and which actions require an independently authorised decision. Restricting unnecessary data access and separating reading from consequential actions can reduce impact. A supplier’s remediation of one flaw is not an enduring guarantee that every possible prompt-injection technique has been solved.

THE NUMBERS / WHAT THEY ACTUALLY MEASURE

Large figures need a denominator and a date.

The sources are the FBI’s Bybit attribution, Cloudflare’s Q3 2025 report, and the UK government’s 2025 skills study. None establishes that employment pressure caused professionals to commit ransomware offences, or that the largest reported attack is the right capacity target for every business.

Use figures to ask a better operational question. For payment approval, who can independently validate the destination? For service availability, what demand or interruption can the business tolerate? For staffing, which essential capabilities are missing and what coverage is actually available?

FIVE QUESTIONS FOR THE BOARD

Move from confidence to evidence.

• What do we trust? Identify the suppliers, people, integrations, and automated workflows carrying authority, and the business owner of each.
• Who or what can act on our data? Include non-human identities and assistants, their permissions, and the route to change them.
• How would misuse become visible? Connect identity, data-access, and application records to business context; do not assume every malicious action will look technically unusual.
• What changes during a crisis? Make emergency access temporary and auditable, with independently checked payment and approval decisions.
• What does assurance actually demonstrate? Identify the process and period examined, unresolved exceptions, and the operational test still needed.

Compliance evidence remains useful. Its limits should be explicit: the scope of an audit, the date it covers, and the relationships it did not examine. Dismissing all compliance as yesterday’s paperwork would be as unhelpful as treating it as proof of continuous security.

READING THE OUTLOOK IN OCTOBER 2026

Keep a forecast separate from an observed event.

Further abuse of normal connections, support requests, and automated work was a plausible concern entering 2026. That is a planning hypothesis, not proof that every new incident followed the same pattern. Keep an incident account, a research demonstration, a measurement, and a forecast visibly distinct.

For the next management review, select one consequential trust relationship and trace its lifecycle: business purpose, approval, permissions, monitoring, emergency use, and termination. Look for a point where everyone assumes somebody else owns the decision.

The lasting insight from 2025 is not that trust must disappear. It is that delegated authority needs an owner, an observable purpose, and a practical way to withdraw it. Those are operating requirements a business can examine today.