BRENNEN

Industries / Technology / Software and SaaS

Security

Keeping software, and the data it holds, safe from attackers: finding flaws, fixing them, getting the fixes to users and catching break-ins.

7

Milestones achieved, from 1988 to 2025.

3

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

In depth

The same ground is charted in more depth, as an industry of its own.

The road

AchievedWatch forOpen

Achieved

  1. 1988

    Worm prompts a response team

    On 2 November 1988 a self-copying program, called a worm, spread across the early internet. Lawrence Livermore National Laboratory puts the damage at about 6,000 of the roughly 60,000 connected computers within hours. Its author, Robert Morris, a Cornell graduate student, became the first person convicted under the US Computer Fraud and Abuse Act of 1986, by the laboratory's account. Afterwards DARPA, the US defense research agency, set up the CERT Coordination Center at Carnegie Mellon University's Software Engineering Institute "to improve communication during emergencies".

    LessonComputers run by many separate owners need a standing place to coordinate when one attack hits them all.

    st.llnl.govdiscmaster.textfiles.com

    Credit

    DARPA; the Software Engineering Institute

  2. 1999

    One name per flaw

    Security tools each kept their own list of known flaws under their own names, so matching one tool's finding to another's was slow work by hand, as two engineers at MITRE, a US nonprofit, wrote in January 1999. They proposed one shared list. The CVE List (Common Vulnerabilities and Exposures) opened to the public in September 1999 with 321 records, and each flaw now gets one ID, such as CVE-2014-0160. By December 2000, 29 organizations had declared their products compatible with it, by the CVE program's count.

    cve.orgcve.org

    Credit

    David E. Mann and Steven M. Christey, MITRE

  3. 2014

    Heartbleed exposes thin funding

    A flaw in OpenSSL, the free code that encrypts much of the web's traffic, let anyone read pieces of a server's memory. The finders used it to steal their own server's secret keys, user names and passwords. The flaw entered the code in December 2011, shipped in March 2012 and was fixed on 7 April 2014. Two weeks later the Linux Foundation announced the Core Infrastructure Initiative: at least 12 companies would each give $100,000 or more a year for three years to underfunded open-source projects, starting with OpenSSL, as TIME reported.

    LessonCode that much of the internet relies on can be kept by a small, thinly funded team, and someone has to pay for its upkeep.

    heartbleed.comtime.com

    Credit

    engineers at Codenomicon and Neel Mehta of Google, who found the flaw independently; the Linux Foundation

  4. 2017

    A patch notice goes astray

    On 8 March 2017 the US Computer Emergency Readiness Team warned of a flaw in Apache Struts, free software for building websites. Two days later attackers found that Equifax, one of the US consumer reporting agencies that keep people's credit records, still ran the flawed version on its online dispute portal. Equifax had sent the warning to an out-of-date list of administrators, and its own scan a week later missed the portal. From 13 May the attackers took personal data for about 76 days, unseen because an expired digital certificate had stopped a traffic-inspection tool from working. At least 145.5 million people in the US were affected.

    LessonA fix protects nobody until the warning reaches whoever runs each affected system, and monitoring tools need checks of their own.

    gao.gov

    Credit

    the US Government Accountability Office (GAO), whose August 2018 report sets out the sequence

  5. 2021

    Log4Shell stays for years

    On 24 November 2021 a security engineer at Alibaba Cloud reported a flaw in Log4j, a free Java library that thousands of programs use to record what they do. It let a crafted log message run code fetched from another server, and it became public on 9 December. Reviewing it in July 2022, the US Cyber Safety Review Board called it an "endemic vulnerability" that would stay in systems "perhaps a decade or longer". It found that none of the organizations it spoke to had used a software bill of materials (SBOM), a list of the parts inside a program, to find their vulnerable copies.

    LessonA flaw can be fixed at its source in days while copies of the old code stay in use for years.

    cisa.gov

    Credit

    the volunteer maintainers at the Apache Software Foundation, who fixed it; the Cyber Safety Review Board, which reviewed it

  6. 2024

    A backdoor caught by chance

    Andres Freund, a software engineer, noticed that logins over ssh, the standard tool for reaching a server remotely, were using too much processor time on his test systems. He traced it to a backdoor in XZ Utils, a compression library, and reported it on 29 March 2024. The hidden code was in the release tarballs (archive files of source code) of versions 5.6.0 and 5.6.1, created and signed by a co-maintainer, Jia Tan, and not in the project's public git history.

    LessonThe files people install can differ from the source that people review, so release files need a check of their own.

    openwall.comtukaani.org

    Credit

    Andres Freund, who found it; Lasse Collin, the original maintainer, who documented it

  7. 2025

    AI finds and patches flaws

    In 2024 Google said its Big Sleep agent had found an exploitable flaw in SQLite, a widely used database, which the SQLite developers fixed the same day. Google called it the first public example of an AI agent finding such a flaw in widely used real-world software. In August 2025 DARPA announced the final of its two-year AI Cyber Challenge (AIxCC), run with ARPA-H. Seven teams' automated systems worked through 54 million lines of open-source code in 63 challenges. They found 54 of the flaws the organizers had planted and patched 43, and they also found 18 real flaws that nobody had planted. The average cost per task was about $152, by DARPA's figures, and four of the seven systems were released as open source.

    darpa.milgoogleprojectzero.blogspot.com

    Credit

    DARPA and ARPA-H; Team Atlanta (first), Trail of Bits (second), Theori (third)

Watch for

  1. Thousands of flaws, few patched

    Anthropic says its Claude Mythos Preview model, shared from April 2026 with partners in Project Glasswing, flagged 23,019 possible flaws in more than 1,000 open-source projects, 6,202 of them rated high or critical. Six outside security firms, and in a few cases Anthropic, checked 1,752 of the severe ones and found 90.6% to be real. Of the 530 severe flaws reported to maintainers by then, 75 had been patched; Anthropic calls maintainers' time the bottleneck. A separate count by VulnCheck, a vulnerability-data firm, found 1,061 flaws credited to AI-assisted discovery in the first half of 2026, of which 14 were confirmed exploited.

    Achieved whenan outside tracker shows that maintainers have confirmed and fixed most of the severe flaws reported.

    anthropic.comthenextweb.com

    Claimed by

    Anthropic, May 2026

  2. A reviewer for every commit

    OpenAI says its Codex Security agent, in its first 30 days of testing, scanned more than 1.2 million commits (saved changes to code) in outside repositories. It flagged 792 critical and 10,561 high-severity flaws, and 14 CVE IDs were assigned for flaws in open-source projects including OpenSSH, GnuTLS, libssh and PHP.

    Achieved whenan outside party checks a random sample of the flagged findings and publishes how many are real.

    csoonline.com

    Claimed by

    OpenAI, March 2026

  3. A high score, self-run

    Google says its Gemini 4 Carbon model scored 88.2% on SEC-bench Pro, a test in which a model must produce an input that crashes the flawed version of a program but not the patched one. Google ran the test itself, on a 183-task split of the 344 tasks the benchmark lists, without guardrails (the limits placed on what a model will do), and gave the model first to a group of vetted cyber defenders (see the Note "Cyber defenders first").

    Achieved whenthe benchmark's maintainers or another outside party run the model on all 344 tasks and publish the score.

    blog.googlestorage.googleapis.comsec-bench.github.io

    Claimed by

    Google, October 2026

Open

  1. 01

    Release tarballs rebuilt from git

    For 50 widely used projects that publish release tarballs, among them xz, gzip, coreutils, tar, sed and bash, a public recipe that rebuilds each project's latest tarball from its tagged git source. Where the rebuild does not match, list every file that differs and why.

    The brief
    Credit · unclaimed
  2. 02

    Fix-to-install time for AI finds

    A public table that follows 200 flaws found with AI help in 2026 from the day the fix landed in the project's own code to the day a fixed package reached Debian stable, Ubuntu LTS, Fedora and Alpine Linux. Beside them, 200 flaws in the same projects found without AI, as a comparison. Updated monthly.

    The brief
    Credit · unclaimed

Open 01 · the brief

Release tarballs rebuilt from git

Credit · unclaimed

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

What to build

For 50 widely used projects that publish release tarballs, among them xz, gzip, coreutils, tar, sed and bash, a public recipe that rebuilds each project's latest tarball from its tagged git source. Where the rebuild does not match, list every file that differs and why.

Why it is still open

The XZ backdoor sat in release tarballs, not in git. In April 2025 the developer Simon Josefsson wrote that he knew of no "automated or collaborative effort" to confirm that release tarballs come from their git sources. His own recipes cover only projects he released, and he gave up trying to rebuild the XZ 5.8.1 tarball because he could not match the build tools its maintainer had used. Google's OSS Rebuild checks published packages against their source, but only for npm, PyPI and crates.io.

blog.josefsson.orggithub.com

Built with what exists

git; Docker or GNU Guix to pin exact versions of build tools such as Autoconf, Automake, Libtool and Gettext; diffoscope, the Reproducible Builds project's tool for showing how two files differ; an AI coding agent such as GitHub Copilot Fusion or Claude Code to search for the tool versions that reproduce each file (see the Note "Effort, not model").

  1. First step

    Pick five projects with a release this year and try to rebuild each tarball from its git tag, logging every difference.

  2. Done looks like

    For 50 projects, either a byte-for-byte match with the published tarball or a list of every differing file with a stated cause. A second person reruns 10 of the recipes on another machine and gets the same SHA-256 checksums.

Open 02 · the brief

Fix-to-install time for AI finds

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 that follows 200 flaws found with AI help in 2026 from the day the fix landed in the project's own code to the day a fixed package reached Debian stable, Ubuntu LTS, Fedora and Alpine Linux. Beside them, 200 flaws in the same projects found without AI, as a comparison. Updated monthly.

Why it is still open

AI tools now find flaws faster than maintainers fix them, and attackers can read a published fix to learn the flaw. In November 2026 Google's Project Zero said that "patch soak time", the wait it allowed for fixes to reach users before publishing details, "is no longer meaningful", and that it will publish flaw details 60 days after reporting from February 2027 and 30 days from January 2028. Anthropic says patched software "is being rolled out much more quickly", citing vendors' patch counts, not times. The latest large outside measurement found, a study of Debian and Fedora published in 2023, put their median fix times at 20 and 27 days, before AI-found flaws arrived in bulk.

projectzero.googleanthropic.comsail.cs.queensu.ca

Built with what exists

lists of AI-credited flaws from public trackers such as VulnCheck's and from the credit fields in CVE records; OSV.dev, the open vulnerability database, for fix versions and commits; the Debian Security Tracker, the Ubuntu CVE tracker, and Fedora's and Alpine's security data; Python and pandas.

  1. First step

    Take 20 AI-credited CVEs in C libraries that Debian packages, and record the upstream fix date and the date of Debian's fixed version for each.

  2. Done looks like

    400 flaws traced, each date linked to the advisory or package changelog it came from. A second person re-checks 40 rows chosen at random and agrees with at least 38.