BRENNEN

Industries / Technology / Software and SaaS / Cybersecurity

Securing software before release

Designing, writing and building software so that fewer flaws reach the people who use it.

7

Milestones achieved, from 1972 to 2025.

2

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1972

    Security cannot be added later

    In a study for the US Air Force dated October 1972, James P. Anderson found that commercial operating systems of the day had "serious design and implementation flaws" that let a programmer seize control of the machine, and that patching could not repair them: "Unless security is designed into a system from its inception, there is little chance that it can be made secure by retrofit." He proposed the reference monitor, a small, tamper-proof part of a system that checks every attempt by a program to reach data or devices against a list of what it is allowed to do.

    csrc.nist.gov

    Credit

    James P. Anderson, for the US Air Force Electronic Systems Division

  2. 2002

    Security over features

    On 15 January 2002 Bill Gates wrote to all Microsoft staff: "when we face a choice between adding features and resolving security issues, we need to choose security." Microsoft was training all its developers in secure coding and reviewing each product for flaws. In 2004 its Security Development Lifecycle (SDL), a set of required security steps from design to release, became mandatory policy at Microsoft, by the company's account, and Microsoft later published it for others to use.

    theregister.commicrosoft.com

    Credit

    Bill Gates; Microsoft

  3. 2003

    Ten risks for web builders

    OWASP (now the Open Worldwide Application Security Project), a volunteer community, first published its Top 10 list of the most serious risks to web applications in 2003. It was meant to raise awareness, but OWASP notes that organizations have used it as a de facto standard since. The 2025 edition draws on test data for more than 2.8 million applications. Broken access control, in which users can reach data or actions they should not, stays first, and software supply chain failures are third.

    github.comgithub.comtop10.owasp.org

    Credit

    the OWASP Top 10 team and its data contributors

  4. 2019

    Most flaws are memory errors

    In July 2019 Microsoft's Security Response Center reported that about 70% of the flaws it assigns a CVE each year are memory safety issues: mistakes in how programs in C and C++ handle memory, which languages such as Rust and C# rule out by design. It said many of its teams had moved new features to such languages. Governments took up the point: CISA's guidance of January 2025 says makers of existing products written in memory-unsafe languages should publish a memory safety roadmap by the end of 2025.

    microsoft.comcisa.gov

    Credit

    Microsoft Security Response Center; CISA

  5. 2023

    AI help, less secure code

    In a user study, people who wrote security-related code with an AI assistant, based on OpenAI's codex-davinci-002 model, wrote significantly less secure code than those without one, and were more likely to believe their code was secure. Participants who trusted the AI less and reworded their prompts produced fewer flaws. The study was published at ACM CCS, a leading security conference, in November 2023.

    LessonCode that arrives quickly and looks right still needs the same security checks, and confidence is not evidence.

    arxiv.org

    Credit

    Neil Perry, Megha Srivastava, Deepak Kumar and Dan Boneh

  6. 2024

    Security duties in EU law

    The EU's Cyber Resilience Act took effect on 10 December 2024. It requires makers of hardware and software products sold in the EU to ship them "without known exploitable vulnerabilities" and with secure default settings, to keep a software bill of materials (SBOM), a machine-readable list of the parts inside a product, and to provide security updates. From 11 September 2026 makers must warn their national response team and ENISA, the EU cybersecurity agency, of an actively exploited flaw within 24 hours of learning of it; the other duties apply from 11 December 2027. Fines reach €15 million or 2.5% of worldwide annual turnover, whichever is higher.

    eur-lex.europa.eudigital-strategy.ec.europa.eu

    Credit

    the European Parliament and the Council of the EU

  7. 2025

    A worm in npm

    In September 2025 a self-copying program called Shai-Hulud spread through npm, the main registry of JavaScript packages. It entered through stolen maintainer accounts. Once installed, it searched developers' computers for passwords and cloud keys, posted them to a public GitHub repository, and used each victim's npm access to publish infected versions of their other packages. More than 500 packages were affected, by CISA's count. GitHub, which runs npm, removed them and blocked uploads carrying the worm. It then announced that publishing tokens would expire after seven days, that publishing would require two-factor login or trusted publishing (linking a package to the build system that publishes it, with no stored token), and that one-time-code logins would give way to hardware security keys.

    LessonA long-lived publishing token is a key to every package its owner maintains, and a worm can turn each stolen key into the next infection.

    cisa.govgithub.blog

    Credit

    GitHub's npm team; CISA

Watch for

  1. A thousandfold fewer memory flaws

    Google says memory safety flaws fell below 20% of all flaws found in Android in 2025, for the first time. It puts Android's history for C and C++ at close to 1,000 memory safety flaws per million lines of code, and estimates 0.2 per million lines for its Rust code, based on about 5 million lines and one possible flaw caught before release. It also says Rust changes are rolled back about 4 times less often than similar C++ changes and spend about 25% less time in code review.

    Achieved whenan outside party recounts the memory safety flaws in Android's public security bulletins by language and reaches similar figures.

    blog.google

    Claimed by

    Google (Jeff Vander Stoep, Android), November 2025

  2. An agent that rewrites code

    Google DeepMind says its CodeMender agent both patches new flaws and rewrites existing code to remove whole classes of them. It checks its own changes with static analysis, fuzzing (feeding programs large numbers of generated inputs), differential testing and SMT solvers, programs that prove or disprove logical statements about code. Over about six months it sent 72 security fixes to open-source projects as large as 4.5 million lines of code, and people reviewed every patch before submission.

    Achieved whenGoogle or the projects publish the list of accepted fixes, and an outside party confirms that they closed the flaws and were not later reverted.

    deepmind.google

    Claimed by

    Google DeepMind, October 2025

Open

  1. 01

    Memory safety roadmaps, checked

    A public table of the software makers that CISA's guidance covers, starting with the largest makers of operating systems, browsers, databases, network equipment and industrial control software, showing whether each published a memory safety roadmap by the end of 2025. For each roadmap: a link, the products and components it covers, its dates, and whether it meets the points CISA lists, such as putting network-facing code first and leading to a "significant, prioritized reduction" of memory safety flaws. Updated every quarter.

    The brief
    Credit · unclaimed
  2. 02

    Android's memory numbers, recounted

    An outside recount of Google's Android figures from public data. Every flaw in Android's monthly security bulletins from 2019 to 2025 is classed as a memory safety flaw or not, and by the language of the code it was in, using the fix commits the bulletins link in the Android Open Source Project. Flaws in closed-source vendor parts, which have no public fix, are counted separately.

    The brief
    Credit · unclaimed

Open 01 · the brief

Memory safety roadmaps, checked

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public table of the software makers that CISA's guidance covers, starting with the largest makers of operating systems, browsers, databases, network equipment and industrial control software, showing whether each published a memory safety roadmap by the end of 2025. For each roadmap: a link, the products and components it covers, its dates, and whether it meets the points CISA lists, such as putting network-facing code first and leading to a "significant, prioritized reduction" of memory safety flaws. Updated every quarter.

Why it is still open

CISA asked makers to publish roadmaps by the end of 2025, and its guidance names no register of those published. Searches in November 2026 found guides to writing a roadmap and no list of who has published one.

cisa.gov

Built with what exists

makers' security and trust-center pages; the signer list of CISA's Secure by Design pledge; an AI research agent to find and summarize candidate documents, with a person checking each; Python to check that links still work.

  1. First step

    List 20 large makers of network-facing products written mainly in C or C++, and search each one's security pages for a memory safety roadmap.

  2. Done looks like

    100 makers covered, each with a yes or no and a link. Each roadmap scored against CISA's points by two people working separately, with their rate of agreement published. Changes logged each quarter.

Open 02 · the brief

Android's memory numbers, recounted

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

An outside recount of Google's Android figures from public data. Every flaw in Android's monthly security bulletins from 2019 to 2025 is classed as a memory safety flaw or not, and by the language of the code it was in, using the fix commits the bulletins link in the Android Open Source Project. Flaws in closed-source vendor parts, which have no public fix, are counted separately.

Why it is still open

Google's figures, including the thousandfold lower flaw density it reports for Rust, come from Google's own analysis of its own data. Searches in November 2026 found coverage that repeats Google's numbers and no independent recount.

blog.google

Built with what exists

the Android Security Bulletins at source.android.com; the Android Open Source Project's git repositories; MITRE's CWE list of weakness types; Python; an AI model to draft a class for each fix, with every memory safety call checked by a person.

  1. First step

    Take the 2025 bulletins, class 50 flaws from their fix commits, and compare the memory safety share with Google's figure of below 20%.

  2. Done looks like

    every bulletin flaw from 2019 to 2025 classed, each with the commit used. A yearly share of memory safety flaws set beside Google's. A second person checks 100 random rows and agrees with at least 90.