Briefing Old systems and supplier trouble
Have your inherited code checked for holes
Finding security flaws in code has become far easier, so a review is a sensible first step for any system you took over, with a person deciding on every fix.
- Published
- Updated
- Reading time
- 5 min read
- Editor
- David Lahuta
In short
- On 20 February 2026 Anthropic opened a limited preview of Claude Code Security, a tool that scans code for security flaws and proposes fixes for a person to review.1
- Two weeks earlier its researchers reported more than 500 serious flaws found and validated in open source code, some undetected for decades.2
- Since April 2026 the same kind of work has run with large technology companies and over 40 further organisations.3
- For a system you took over from another supplier, a security review of its code is now a sensible first step. The tool finds and proposes, a person reviews, tests and decides on every fix.1
What was released in February 2026?
On 20 February 2026 Anthropic opened Claude Code Security as a limited research preview.1 It scans code for security flaws, called vulnerabilities, and suggests fixes for human review.1 It was open to companies on its Enterprise and Team plans and, free of charge, to maintainers of open source projects.1
The usual automated checks match code against known patterns, so they catch common issues such as exposed passwords or outdated encryption, but often miss flaws in business logic or broken access control.1 The new tool, by the lab's account, reads the code the way a security researcher would, following how its parts interact and how data moves through it.1
Each finding is checked again to remove false alarms and rated for severity and confidence.1 Nothing is applied without human approval, the tool identifies problems and suggests solutions, and developers make the call.1
What does the lab say it found?
| Date | What the lab published | What it reported |
|---|---|---|
| 5 February 2026 | Its Frontier Red Team's results with the model Claude Opus 4.62 | More than 500 serious flaws found and validated in open source code, some in projects tested with automated tools for years and undetected for decades2 |
| 20 February 2026 | Claude Code Security, a limited research preview1 | Findings and suggested fixes, with nothing applied without human approval1 |
| 7 April 2026 | Project Glasswing, with Amazon Web Services, Apple, Google, Microsoft, the Linux Foundation and others3 | Thousands of serious flaws found by an unreleased model, Claude Mythos Preview, including in every major operating system and web browser3 |
In the February research every flaw was validated before it was reported, and in the first round the lab's own security researchers validated each one and wrote the fixes by hand.2
In April the lab showed how long flaws can hide. A flaw in OpenBSD, an operating system known for its security, was 27 years old.3 One in FFmpeg, video software, was 16 years old, in a line that automated testing tools had hit five million times without catching it.3 Both are now fixed.3
Why does this matter for a system you took over?
A system inherited from another supplier often comes with thin documentation and nobody on your side who knows its code. It also rests on open source parts, often maintained by small teams or volunteers.2
Software has always contained bugs, and some let attackers take over systems, disrupt operations or steal data.3 The cost, effort and expertise needed to find and exploit such flaws have dropped sharply.3 The same capabilities that help defenders could help attackers.1 The lab's conclusion is that defenders who move quickly can find the same weaknesses and fix them first.1
What should a review of an inherited system cover?
Agree in writing before it starts what you will receive.
Before you decide
- A list of what the system consists of, including the open source parts and their versions.
- A scan of the code for known patterns and for flaws in business logic and access control.1
- Every finding confirmed by a person and rated by severity, with false alarms removed.12
- Proposed fixes reviewed, tested and approved by a person before anything changes in the running system.
- A plain report of what was found, how serious it is, what was fixed and what remains.
- An agreement on who watches for new flaws in the system and its parts afterwards.
What does a person still decide?
The tool speeds up the search, the judgement stays with people. The lab notes that such flaws often involve details that are hard to judge from the code alone.1
| Step | What the tool does | What a person does |
|---|---|---|
| Find | Scans the code and lists possible flaws1 | Decides which systems and which parts are scanned |
| Confirm | Checks each finding again, rates severity and confidence1 | Confirms whether the flaw is real in your system |
| Fix | Proposes a fix1 | Reviews and tests it, decides whether and when it goes live |
| Report | Nothing, this step is not the tool's | Decides what to tell management, customers or an authority |
Act now or wait?
Act now if the system holds customer or personal data, can be reached from the internet, or its original supplier is gone, and also if NIS2 applies to you or you sell software of your own.
Waiting is reasonable if the system will be switched off within months and is not reachable from outside. Even then, list its parts.
The tools are new, and the strongest model the lab describes is not planned for general availability.3 The lab expects a significant share of the world's code to be scanned by AI soon.1 A review does not depend on one tool. What matters is that a person you trust confirms the findings and answers for every fix.
What this means for a company in Europe
For companies covered by NIS2, the EU directive on cybersecurity, such a review is part of a duty. Their measures must include security in acquiring, developing and maintaining systems, including handling vulnerabilities.4 They must also take into account the quality of their suppliers' products and their secure development practices.4 Management approves these measures, oversees them and can be held liable.4 National laws had to apply from 18 October 2024.4
The Cyber Resilience Act adds duties for manufacturers, meaning whoever develops software or devices with software and markets them under its own name, paid or free.5 Since 11 September 2026 a manufacturer must report an actively exploited vulnerability, one that an attacker has reliably been shown to use, to its national cybersecurity team and to ENISA, the EU cybersecurity agency.5 An early warning is due within 24 hours and a fuller notification within 72 hours.5 Most other duties apply from 11 December 2027.5
Next step
If you have taken over a system and would like to know what a review of its code would involve, we are glad to scope one for that system with you in a short meeting. What you do with the result is entirely your decision.
Sources
- 1. Anthropic, Making frontier cybersecurity capabilities available to defenders (20 February 2026). Read on 17 November 2026
- 2. Anthropic Frontier Red Team, Evaluating and mitigating the growing risk of LLM-discovered 0-days (5 February 2026). Read on 17 November 2026
- 3. Anthropic, Project Glasswing, Securing critical software for the AI era (7 April 2026, updated 6 October 2026). Read on 17 November 2026
- 4. Directive (EU) 2022/2555 (NIS2), Articles 20, 21 and 41, Official Journal of the EU. Read on 17 November 2026
- 5. Regulation (EU) 2024/2847 (Cyber Resilience Act), Articles 3, 14 and 71, Official Journal of the EU. Read on 17 November 2026
Change log
- Published.