Did you know that we have LinkedIn, Instagram and X accounts
that you can follow?
Hi {{first_name|everyone}},
Anthropic’s models found so many potential software bugs that its own human reviewers can’t keep up, and some open-source maintainers now just want the raw findings sent over. Then, we’ve got a backwards way of making AI art, a few handy tools, an imaginary band made entirely by AI, and a quick share from our friends on the job market for forward deployed engineers. Okay, time to dig in.
🤿 DEEP DIVE
Claude Found 29,000 Potential Bugs
The surprising part is how many of those findings look real.
Apparently, AI has gotten so good at finding software vulnerabilities that Anthropic is struggling to keep up with its own discoveries.
Over the last six months, the company used its latest models to scan some of the world's most important open-source software projects.
The result? More than 29,000 potential vulnerabilities.
But here's the problem. Its human security experts only managed to review around 6,000 of them.
So there are thousands of findings sitting outside the normal review process, and some open-source maintainers have started asking Anthropic to just send everything over, even the stuff nobody has checked yet.
And honestly, the early results help explain why.
The reports aren't just AI guesses anymore
Anthropic says that on CyberGym, a benchmark for finding software vulnerabilities, language models went from identifying less than 20% of vulnerabilities at the beginning of last year to more than 85% this year.
That's quite a jump.
But benchmarks are one thing. What happens when actual software maintainers receive these reports?
Well, Anthropic tested an early version of its OSS Scanner by having expert security testers review 97 high- and critical-severity findings across 48 projects.
Of those, 85 met Anthropic's standards for verified vulnerability disclosures. Another 11 were real issues but duplicated existing findings.
Only one was actually invalid.
That's a pretty interesting result for reports generated entirely by AI.
And the feedback from maintainers was similarly encouraging.
The team behind wolfSSL received 74 reports, with all but two considered valid. Five became officially recorded vulnerabilities, known as CVEs.
Meanwhile, curl maintainer Daniel Stenberg said the scanner helped uncover multiple issues worth fixing, including one of the worst curl vulnerabilities reported in recent years.
So why not just send everything?
Well, that's essentially what Anthropic is experimenting with.
Its new OSS Scanner gives participating open-source projects regular, free vulnerability scans using its strongest models, including Claude Mythos.
Each report includes an explanation, a way to reproduce the problem, and a suggested patch when available.
And unlike Anthropic's traditional disclosure process, these reports go straight to maintainers without human verification.
There's still a trade-off here. Reports can contain incorrect severity ratings, duplicate findings, or misunderstand how a particular project handles security.
So the scanner isn't replacing human judgment. It's giving projects that have the resources to investigate findings a way to receive them much faster.
And that seems to be the real shift.
AI isn't just getting better at spotting security problems. It's reaching the point where some maintainers would rather receive the raw findings and decide what to fix themselves than wait for every report to pass through a human review queue.
The quick take
29,000+ potential vulnerabilities were discovered by Anthropic's models over six months, but only around 6,000 were manually reviewed.
Only 1 of 97 tested findings was invalid, although 11 were duplicates and 85 met Anthropic's disclosure standards.
Real maintainers are seeing useful results, including wolfSSL, PostgreSQL, OpenSSL, and curl.
OSS Scanner sends unreviewed AI findings directly to participating projects, with explanations and suggested fixes where available.
The big shift is that human verification is becoming a bottleneck as AI vulnerability discovery improves.
🤝 FROM OUR FRIENDS
A quick share we thought you might find useful
Sponsored ———————————————————————————————————————————————————————
AI's next bottleneck is deployment
New models are easy to demo. Making them work inside real customer operations is not.
That is where forward deployed engineers come in: bridging code, customer context, and production outcomes.
The free State of FDE Jobs 2026 report maps the emerging labor market around this work.
🔥 WHAT’S THE BUZZ
What people are building, sharing, and talking about
It is a wonderfully backwards way to make AI art. Instead of asking a diffusion model to generate a painting directly, a language model has to translate its internal understanding of a painter or style into thousands of explicit drawing instructions.
The broader idea is more interesting than the paintings themselves. A capable language model does not necessarily need a purpose-built tool for every creative job. If it understands the underlying primitives well enough, it can sometimes construct the output from scratch using ordinary software.
⚒ TOOL SNAPSHOTS
Futuristic tools within AI, no-code, and productivity
Set up recurring tasks once and let them keep running.
Why it’s useful: Handy for teams that need regular reports, source monitoring, or tracker updates without having to prompt an AI every time.
🎯 HireKey
Get a better sense of which jobs fit and how to land them.
Why it’s useful: Helps tech job seekers sharpen their applications and practice interviews out loud, with feedback informed by real recruiting experience.
Keep up with your audience without sounding like a bot.
Why it’s useful: Useful for creators juggling comments and DMs across platforms who want replies that sound like them, remember past conversations, and still get their approval before sending.
🎁REFERRAL REWARDS
If TIW has been useful, invite one friend or coworker who’d actually use AI/automation at work.
1 referral → get the TIW No-Code Workflows PDF (25 plug-and-play workflows)
10 referrals → get a full course free (your choice)
Your referral link: {{rp_refer_url}}
Tip: send it to a friend who’s AI-curious, a coworker who wants to move faster, or anyone trying to automate repetitive tasks.
🤖 AI MADE THIS
Interesting and inspiring creations made with AI
An imaginary band made entirely out of surreal AI characters
AI used
Midjourney + Nano Banana Pro for imagery and visual concepts
Google Veo 3 for animation/cinematic motion
Suno for original music from Crystal Trimble's human-written lyrics and creative direction
ℹ️ ABOUT US
The Intelligent Worker helps you to be more productive at work with AI, automation, no-code, and other technologies.
We like real, practical, and tangible use-cases and hate hand-wavy, theoretical, and abstract concepts that don’t drive real-world outcomes.
Our mission is to empower individuals, boost their productivity, and future-proof their careers.
We read all your comments - please provide your feedback!
Did you like today's email?
What more do you want to see in this newsletter?
