Data Breaches in 2026: 10 Recent Attacks, Five Involving AI, and What Would Have Stopped Them

The Short Answer

What keeps happening: in ten widely covered breaches from July to November 2026, six turned on a key, token or login in the wrong hands, three came through a third party or a dependency, and five involved AI, either as the system that leaked or as the attacker’s tool. Almost none needed a new kind of attack.

What would have stopped them: each cause has a control that targets it. An audit finds what is switched off or left behind: a database without access rules, a key in a build file, records from 1978. A review asks what a stolen key, a fooled agent or a third-party app could reach. Monitoring notices the bulk download or the spending spike while it is happening.

How rare those controls are: in the UK government’s Cyber Security Breaches Survey 2025/2026, 43% of businesses reported a breach or attack in the last year, rising to 69% of large ones, yet only 18% ran a vulnerability audit and 32% used security monitoring tools.

Ten Breaches and Attacks, July to November 2026

Chosen because they were widely reported and their cause was published, weighted to UK organisations and to AI. Listed in order of disclosure. “Control” is the safeguard aimed at that cause; it is my reading, not the company’s.

What the Ten Have in Common

1. Keys, tokens and logins did most of the damage

Beacon (a cloud key), ChainDrop (a maintainer’s account), Master of Malt (a vendor’s application key), ASOS (an employee’s login), OpenCode (stored AI keys) and Chainalysis (passwords and tokens). The attacker did not break the software; they used a credential it trusted. The two questions that limit the damage are what each credential can reach and how long it stays valid. On the pharmaceutical platform described below, access tokens lasted 7 days before the review and 8 hours after it.

2. Third parties and dependencies are part of your system

Verizon’s 2026 Data Breach Investigations Report found a third party involved in 48% of breaches, up 60% on the year before. The DSIT survey found only 15% of UK businesses reviewed the risks from their immediate suppliers. Master of Malt’s shoppers lost their data through an app vendor’s key, and ChainDrop reached companies that had never heard of the package it came in.

3. AI raised the stakes on both sides

AI systems are now something to breach. IBM’s 2026 Cost of a Data Breach study found roughly one in five organisations reported a breach of an AI model or application, and 92% of those lacked proper AI access controls. UpGuard’s Supabase research shows the same failure in apps built with AI coding tools: tables created by an agent start without access rules, a gap explained in why AI-built apps leak data. Secrets leak the same way: my study of 53 public AI-built repositories found a committed .env file in at least 19% of them.

AI is also the attacker’s tool. Chainalysis and WorkOS both describe AI-assisted intrusions, IBM recorded a 56% rise in AI-driven attacks, and 22% of UK organisations reported AI-generated attacks. Automated attackers find a forgotten internal tool or a stale key faster than a person does, which shortens the time a gap can stay open unnoticed.

4. Old data and forgotten systems

The Legal Aid Agency confirmed about 250,000 stolen records from before 2007, the earliest from 1978. OpenCode’s keys were taken from an older system; WorkOS’s exposure was an internal sales tool. Retention policies and an inventory of what still runs are not exciting controls, but data that no longer exists cannot be stolen.

Audit, Review and Monitoring: What Each One Catches

None of the three replaces the others. An audit is a snapshot, so it misses what changes the week after. A review prevents structural problems but cannot see an attack in progress. Monitoring without the first two produces alerts about problems that were avoidable. UK guidance asks for all three: the DSIT Code of Practice for the Cyber Security of AI asks for security testing before release and says system operators “shall log system and user actions” (principle 12).

A Worked Example: All Three on One AI Platform

I spent seven weeks as fractional CTO and security architect on a multi-tenant AI platform for pharmaceutical manufacturers, heading into pilots with large pharma companies. It worked, but it had several of the gaps in the ten incidents above.

• Audit. 26 pass/fail security checks found 2 critical, 4 high and 3 medium issues. One critical finding was the Supabase pattern in a different stack: one tenant could read another tenant’s documents.
• Review. Access tokens lasted 7 days, so a stolen token meant a week of access to manufacturing data; they now last 8 hours. Six Kubernetes deployments stopped running as root, and admin access was limited to named accounts.
• Monitoring. The platform had no central logging, alerting or health checks. It now has 8 CloudWatch alarms, an audit trail covering 12+ critical actions kept append-only for 365 days, and every AI call traced to its tenant, user and request. Guardrails for prompt injection and data leaks can block or only log, so they can be tuned on real traffic instead of being switched off.

Both critical and all four high findings were fixed, all 22 contracted deliverables were completed with evidence, and the platform was approved for enterprise pilots. It was not a penetration test and it is not a certification; the case study lists what it does not claim.

What to Check This Month

1. List every key and token that can read customer data, and how long each stays valid.
2. Scan the files you ship (JavaScript bundles, mobile builds, containers) for secrets, not only the repository.
3. Check every database table for row-level security or an equivalent rule. The Supabase security checklist covers it step by step.
4. List third-party apps and vendors with access to customer data, and remove what is not used.
5. Require phishing-resistant sign-in on every SaaS tool that holds customer data.
6. Write down what each AI agent can do, and make it act with the user’s permissions, not its own (AI agent authorization).
7. Set alerts on bulk reads, logins from new locations and AI spend that jumps.
8. Log every model call with tenant, user and request ID (what to log).
9. Delete data you no longer need, starting with the oldest.
10. Find the internal tools and legacy systems nobody owns, and either secure them or switch them off.

The Numbers Behind This

• 43% of UK businesses and 28% of charities reported a breach or attack in the last 12 months; 65% of medium and 69% of large businesses (DSIT, April 2026).
• In the same survey, 30% of businesses ran a cyber risk assessment, 18% a vulnerability audit and 13% a penetration test. Of businesses using, adopting or considering AI, about a quarter (24%) had security practices in place to manage its risks.
• £3.13 million average cost of a breach for UK organisations; $4.99 million globally, a record (IBM UK, IBM, July 2026).
• 80 nationally significant incidents handled by the NCSC, out of 337 that needed its support and 1,831 reports, in the year to 31 August 2026 (NCSC Annual Review 2026).
• 48% of breaches involved a third party (Verizon DBIR 2026).

Regulation is moving the same way. The Cyber Security and Resilience Bill, still before Parliament in November 2026, would bring managed service providers and data centres into scope and require significant incidents to be reported within 24 hours, with a full report within 72.

Where an Outside Review Fits

This is the work I do. An AI security and architecture review covers the audit and review columns above for an LLM, RAG or agent system, and sets up the monitoring it needs: from £2,000 over 1-4 weeks, fixed in writing before work starts. For an app built with Lovable, Bolt, Cursor or another AI tool, the pre-launch app audit starts from £1,000. If your customer needs a penetration test certificate as well, this guide names the UK firms that do them.

Frequently Asked Questions

What were the biggest data breaches and cyber attacks of 2026?

Widely covered incidents between July and November 2026 include the Beacon charity CRM breach and the ASOS and Master of Malt customer data breaches in the UK, the Legal Aid Agency confirming that stolen records went back to 1978, the ChainDrop worm across more than 1,300 npm packages, UpGuard finding 16,326 Supabase databases with readable tables, and AI-related incidents at OpenAI, OpenCode, Chainalysis and WorkOS. The NCSC handled 80 nationally significant incidents in the year to 31 August 2026.

What causes most data breaches in 2026?

Stolen or exposed credentials and keys, third parties, and misconfiguration. Verizon’s 2026 Data Breach Investigations Report found a third party involved in 48% of breaches and vulnerability exploitation as the first step in 31%. In the ten incidents on this page, six turned on a key, token or login in the wrong hands, and three came through a third party or a dependency.

How much does a data breach cost a UK business?

IBM’s 2026 Cost of a Data Breach study puts the average for UK organisations at £3.13 million, down from the previous year, against a global average of $4.99 million. IBM’s sample is mostly larger organisations; for a small company the cost is usually measured in lost customers, regulator time and a stalled enterprise deal rather than a single bill.

What is the difference between a security audit, a security review and security monitoring?

An audit checks a system against a list at a point in time: is row-level security on, are secrets out of the build, is old data deleted. A review reads the design and the code to ask whether a failure stays contained: what a stolen key can reach, what an agent can do, what a third-party app can read. Monitoring watches the running system all the time and raises an alert when something unusual happens. An audit and a review prevent; monitoring catches what got through.

How often should a company run a security audit?

At least once a year, and before any change that raises the stakes: a launch, an enterprise pilot, a new AI feature or agent, or a new supplier with access to customer data. Monitoring should run all the time. Only 18% of UK businesses ran a vulnerability audit in the last year, and 32% used security monitoring tools, according to the government’s Cyber Security Breaches Survey 2025/2026.

Are AI systems more likely to be breached?

They add new ways in. IBM’s 2026 study found roughly one in five organisations reported a breach of an AI model or application, and 92% of those lacked proper AI access controls. Attackers are using AI too: IBM recorded a 56% rise in AI-driven attacks, and 22% of UK organisations reported AI-generated attacks. Most AI breaches still come down to familiar failures: an over-permissioned agent, an exposed key, or a database without access rules.