AI safety open source gets $1B pledge to harden networks
Anthropic's Cyber Mission pairs its top Claude models with security firms guarding utilities, and offers free AI bug scans to key open-source projects.
Anthropic launched the Anthropic Cyber Mission in October with two parts: an infrastructure defense program and a free open-source scanner.
Eleven partners, including CrowdStrike, Hitachi and Booz Allen, will use Claude to help utilities and plants fix flaws.
OSS Scanner sends free, AI-written vulnerability reports to eligible open-source projects without human review first.
Anthropic says its models flagged more than 29,000 candidate flaws in six months but people could check only about 6,000.
Quick read · 1 min
Anthropic has launched the Anthropic Cyber Mission, a two-part security push. The Critical Infrastructure Defense Program puts its strongest Claude models and on-site engineers alongside eleven firms that protect utilities and factories, including CrowdStrike, Hitachi, Dragos and Booz Allen.
The second part, OSS Scanner, gives eligible open-source projects free, regular AI security scans. Anthropic says its models flagged more than 29,000 possible flaws in six months, far more than its staff could review, so reports now go straight to maintainers with test cases and suggested fixes.
No dollar figure was announced; an earlier version of our story was wrong about a billion-dollar pledge.
Reports skip human review, so some severity ratings may be off.
Anthropic expects AI to favor defenders within two years.
For ordinary users, the payoff is fewer unpatched holes in everyday software, as long as you keep installing updates.
Anthropic’s AI models have found so many security holes in the software the world runs on that its own human experts can’t keep up with checking them. So the company is now handing much of that work back to the people who maintain the code, for free, and sending its strongest models and engineers to help the firms that guard power grids and water systems.
This matters even if you have never heard of open-source software. Free, community-built code sits inside your phone, your router, your bank’s website and the systems that keep your lights and taps on. When AI makes it cheaper for attackers to find flaws in that code, defenders need to find and fix them first. Anthropic says attackers hold the upper hand for roughly the next two years, and this is its plan to shorten that window.
01
A correction on the money
An earlier version of this article described a billion-dollar pledge and a chip packaging deal. Neither appears in Anthropic’s announcement. The program, called the Anthropic Cyber Mission, did not come with a dollar figure. Anthropic says the free scanning is paid for by its Defender Advantage Fund, which it set up in August, and it has also given money to the Python Software Foundation, the Apache Software Foundation, and the Alpha-Omega and OpenSSF security efforts run through the Linux Foundation. Early coverage also noted that Anthropic has not said whether its partners get free model access or who pays the computing bill.
02
The Critical Infrastructure Defense Program
The first half of the Cyber Mission targets operational technology, the computers and controllers that run factories, electrical substations and water plants. That equipment is often built to work for decades and can’t simply be switched off for an update, so known weaknesses can linger for years.
Instead of going to every utility directly, Anthropic is working through the outside firms those operators already trust for advice on which fixes are safe. Eleven of them signed on as founding partners. Security vendors form the largest group: CrowdStrike, Palo Alto Networks and the industrial specialists Nozomi Networks, Dragos and Insane Cyber. The consulting giants PwC, Deloitte, Booz Allen and Accenture are in too, along with Rockwell Automation and Hitachi, which build and patch the industrial hardware itself.
Anthropic says it will pair its most capable models with engineers placed on site, because many of these organizations lack in-house AI expertise. Several partners already use Claude to fix flaws, and more partners and industries are expected over the coming months. Booz Allen’s Andrew Turner warned that industrial systems are the next big target for attacks carried out by AI on its own, and that the real question is how much control such an attack could gain, and how quickly.
03
OSS Scanner: free AI security reports for open source
The second half is OSS Scanner, a free, opt-in service that runs regular security sweeps of open-source projects using Anthropic’s strongest models, including Claude Mythos. It grew out of a backlog. Over six months, Anthropic’s models flagged more than 29,000 possible vulnerabilities in widely used software. Staff could manually review about 6,000. Many maintainers asked for the rest anyway, unchecked and with suggested patches attached, and nearly 5,000 of those raw reports have already gone out.
Each report comes with a working reproducer, a small test that shows the bug actually happening. Where it can, the model also proposes a patch and traces back through the code’s history to find when the problem first crept in. For OpenSSL, the encryption toolkit behind much of the web, Anton Arapov said that when a working exploit comes included, the fix is “basically job done” for his team.
Not every project qualifies. Anthropic borrowed its rules from OSS-Fuzz, Google’s long-running service that throws random inputs at open-source code to make it crash. That means established projects with a big effect on infrastructure and user security, judged case by case. Maintainers apply by opening a pull request on Anthropic’s GitHub page with a setup file that lets the scan run offline. Shortly after launch, 116 requests had been filed.
04
The catch: reports skip human review
To move faster, OSS Scanner sends its findings straight to maintainers without a person checking them first. Anthropic admits that means some mistakes, such as a bug rated more severe than it really is, or a misunderstanding of how a project is meant to be used.
To measure that, Anthropic had expert penetration testers, the hackers it hires to vet its disclosures, check 97 critical and high-severity findings from an early version spread across 48 projects. They approved 85 for disclosure. Of the other 12, eleven were real bugs that duplicated known issues or other findings, and only one was invalid. That is a hand-picked sample, not a guaranteed accuracy rate.
Real-world reports have been mostly positive. The team behind wolfSSL, an encryption library used in small devices, said all but two of the 74 reports it received were valid and five became official CVE entries, the public catalog of security flaws. The curl project credited an Anthropic researcher using the scanner with finding a serious HTTP/2 upload bug, which was fixed in a later release.
Because the reports are unverified, Anthropic is not putting its usual 90-day public disclosure clock on them. If a person later confirms one, the normal clock starts from that moment. Projects can pause the reports or leave the program at any time.
You don’t need to do anything. The benefit, if this works, is quieter: fewer holes left open in the software inside your devices and in the systems that supply power and water. Faster fixes upstream reduce the chance that a flaw ends up in the next big data breach or outage.
There are limits. A patch in the original project doesn’t help much if the app or company you rely on runs an old version and never updates. And fixes for industrial equipment can take a long time to roll out safely. For your own devices, the best move hasn’t changed: install updates for your phone, computer and router promptly, because that’s how these upstream fixes eventually reach you.
06
Quick answers
Is OSS Scanner really free?
Yes, for projects that qualify. Anthropic says its Defender Advantage Fund covers the cost. Under the terms that apply to it, Anthropic may use the inputs and outputs for model training.
Did Anthropic pledge a billion dollars?
No. The announcement doesn’t give a dollar amount for the Cyber Mission, and our earlier report of such a pledge was wrong.
Why is Anthropic doing this now?
Anthropic says AI has made finding vulnerabilities easy while checking and fixing them is still slow, human work. It expects AI to favor defenders within two years.
An open-source framework that hands most of a break-in to AI agents is still circulating. Here's who built it, what it hit, and what the FBI says to do.
Some Android beta testers can now skip Google Drive and save chats on WhatsApp's servers, locked by default. Here's how it works and how to protect yours today.
6 min read
We use cookies to understand how readers use Talk With Tech, so we can make it better. Is that OK? Privacy policy
The Daily Brief
Today's biggest tech stories, in 5 minutes
Every morning, the news that matters from AI, phones, apps and the people shaping tech. Explained in plain English. Free.
One email a day. No spam, unsubscribe anytime. Privacy policy