The Exploit Is No Longer in the Code. It’s in the Conversation.

You used to have time to breathe. For a decade, maintaining an open-source project meant fielding a security disclosure every few months. You’d investigate, patch, and ship. Today, you wake up to 40 in a single month. You are drowning, and the water is rising faster than you can tread.

This isn’t a hypothetical. This is the current reality for open-source maintainers. The distance between a vague rumor of a bug and a fully weaponized exploit has collapsed to zero. We thought AI would debug the world. Instead, it weaponized our own conversations.

The old rules of cybersecurity were simple: secure the code, patch the vulnerabilities, keep the bad actors out of the repository. But Large Language Models have obliterated that paradigm. The exploit is no longer just in the code—it’s in the communication.

A maintainer leaves a passing comment in a commit message. A researcher publishes a talk abstract with three vague sentences about a potential flaw. A developer tweets about a weird edge case they found. In the past, these were just whispers. Today, an attacker feeds those whispers into an LLM, and the AI reverse-engineers a working proof-of-concept in minutes.

When the rumor becomes the exploit, your attack surface isn’t your codebase—it’s the entire internet.

I saw this firsthand in the trenches. An rclone maintainer noted that they received 20 security disclosures through GitHub in their first 10 years. Last month alone, they dealt with over 40. Another developer mentioned going from a 3-sentence talk abstract straight to a critical exploit. The AI doesn’t need to understand the deep architecture; it just pattern-matches the abstract concept into a functional attack.

This creates a vicious, suffocating irony. The same AI tools that make it easier to find and fix bugs also make it infinitely easier to mass-produce exploits. And while maintainers are drowning in AI-generated attack vectors, corporate incentives haven’t changed. Your boss doesn’t care that you’re triaging 40 threats a week. They want speed.

As one frustrated developer put it: their boss will give a 30-minute lecture on why they don’t need to solve a bug that Claude fixed in 5 minutes, all while pushing for faster shipping. The organizational will to actually secure infrastructure is eroding, replaced by a mad dash to ship features built on a foundation of sand.

We are asking a handful of underpaid, overworked maintainers to hold back an ocean of automated, AI-generated threats. The trust we place in critical digital infrastructure is quietly rotting from the inside out.

You can’t patch human burnout, and you can’t scale trust. But we’re trying to do both anyway.

Neutrality in the face of this is complicity. This isn’t just a technical problem; it’s an industrial crisis. We have handed every script kiddie and nation-state actor a machine that turns human communication into malware. Until organizations stop rewarding shipping speed over security, and until we accept that our words are now attack vectors, the breaches will only get faster, bigger, and more devastating.

The code was never the weakest link. We were.