AI Engineer World's Fair 2026
The AI bugpocalypse is here. Now what?
Loading slides…
Read the talk
The AI bugpocalypse is here. Now what?
AI is expanding both software production and vulnerability discovery. Jack Cable explains how memory safety, contextual guardrails and defender access can turn that acceleration toward more resilient systems.
From a talk by Jack Cable
More vulnerabilities in the software everyone depends on
What happens when models can discover and exploit vulnerabilities faster, while simultaneously writing more of the software that needs defending? The problem begins in the open-source libraries beneath everyday applications. More discoveries in those shared dependencies mean more exposure across the systems that rely on them—the vulnerability surge Jack Cable calls the AI bugpocalypse.
Cable approaches the problem from both discovery and prevention. He co-founded Corridor to secure AI coding roughly 18 months before this talk, after working as a senior technical advisor at CISA on helping software companies build products that are secure by design. His background also includes reaching HackerOne’s top hundred hackers in high school and studying computer science at Stanford. That experience makes recurring, preventable vulnerability classes central to his response.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Coding and attacks become more autonomous together
The growth of Cursor and Claude Code makes the first change visible: AI coding tools are becoming part of ordinary development. At the same time, frontier models are improving at finding and exploiting vulnerabilities. These changes compound one another. More generated software expands the attack surface, while better discovery makes that surface easier to search.
Cable cites the prior year’s Stack Overflow survey to establish the scale of adoption. Its 84% figure covers respondents using or planning to use AI tools in development, rather than current coding-agent users alone. Cable also recalls 30–40% of companies encouraging AI coding assistance, although that figure is not established by the inspected survey results. His expectation of still higher adoption in the current year is a forecast, not another survey measurement.
The more consequential shift is how developers delegate. Autocomplete leaves a person continuously selecting the next change. Synchronous work in an editor gives an agent a larger task but keeps the developer nearby. Cable describes his team moving beyond that pattern: launching multiple agents from Slack and letting them work in the background. More code can now be produced without a person following every step.
Attackers gain a similar expansion in scope. Cable says models can already outperform his own vulnerability-discovery work, despite his experience reporting hundreds of flaws. He points to an Anthropic chart comparing Mythos with other models, then widens the concern from discovery to exploitation and autonomous attack chains. A defender has to account for automation across the attack process, not just a faster search for suspicious lines of code.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
A new discovery need not be a new vulnerability class
That broader attack capability brings Cable back to Secure by Design: make systems resilient to the attacks they will face, rather than depend entirely on finding each flaw first. Simple vulnerability classes were already recurring before AI. Models make their discovery easier, so defenders need to use the same capabilities to harden the underlying systems.
There is a useful distinction between finding a previously unknown flaw in a particular file and discovering a new kind of flaw. Cable’s assessment is that nearly all the vulnerabilities frontier models are finding belong to familiar classes. That creates a defensive opportunity: a control that prevents an entire class can protect against discoveries that have not happened yet. The question becomes how to apply AI to strengthen those controls as software production changes.
Cable traces this approach to a government Secure by Design paper he recalls publishing in March 2023, when coding LLMs were largely autocomplete tools. The premise does not require perfect security. It requires choosing designs that are substantially more resilient to known failure modes. On the MITRE list he presents, drawn from CISA’s Known Exploited Vulnerabilities Catalog, the leading classes are problems developers have known about—and known how to prevent at scale—for decades.
Buffer overflows appear second on the displayed list. Cable recalls documented examples going back more than 30 years and identifies them as the same established kinds of flaws that models such as Mythos now find. Rust and Go offer a different starting point from C and C++: language and runtime guarantees prevent ordinary code from performing many unsafe memory operations. Rust’s guarantee needs a boundary, however: unsafe operations leave additional correctness obligations with the programmer. Memory safety is a powerful class-level defense, not a claim that every program written in a safer language is secure.
A small Rust example illustrates the mechanism. For a three-byte packet, an attempted read at index 3 produces None rather than reading beyond the buffer:
rust
fn read_byte(packet: &[u8], index: usize) -> Option<u8> {
packet.get(index).copied()
}
fn main() {
let packet = [0x41, 0x42, 0x43];
assert_eq!(read_byte(&packet, 2), Some(0x43));
assert_eq!(read_byte(&packet, 3), None);
}
The caller still has to decide what a missing byte means for the application. But that decision happens without an out-of-bounds memory read. This is the kind of structural constraint Cable wants defenders to exploit.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Safer new code and the case for structural repair
Cable cites CVE-based estimates that memory-safe languages could prevent approximately 60–70% of vulnerabilities in products written in memory-unsafe languages. That is an estimate tied to those products’ vulnerability mix, not a universal prevention rate. He points to adoption at Google, Microsoft and Amazon, along with Rust work in parts of the Linux kernel, as evidence that this is a practical engineering direction.
Android provides the concrete example. Cable emphasizes that the improvement did not require rewriting the entire operating system: Google was writing more new code in memory-safe languages. He describes the chart approximately; Google’s published figures put memory-safety vulnerabilities at 76% of reported Android vulnerabilities in 2019 and 35% in 2022. Those figures cover critical- and high-severity vulnerabilities reported in Android security bulletins. The trend is consistent with safer new code helping, although Google cautions against assuming causation from the correlation alone.
This changes the investment question. Finding and patching individual vulnerabilities remains worthwhile. But repeated discoveries in a foundational library can also justify changing the implementation so the same class stops recurring.
| Investment | What it changes |
|---|---|
| Discover and patch a flaw | Removes a particular vulnerable behavior |
| Write new code in a safer language | Avoids adding many memory-safety flaws |
| Rewrite a critical library in Rust | Changes the memory-safety constraints on that library |
Cable contrasts spending millions on repeated discovery and patching with a one-time rewrite that can keep paying off. The attraction is durability: stronger attacking models do not erase the language’s safety guarantees. Within their supported boundaries, those guarantees continue to constrain which vulnerabilities can exist.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Intelligence does not supply the missing threat model
Safer memory access addresses one family of problems. It does not ensure that a capable model understands the security meaning of the code it writes. Cable recounts a smart-contract incident in which Opus 4.6 allegedly introduced a vulnerability followed by a couple of million dollars being stolen. He uses the account to illustrate a contextual failure: a model can be highly capable and still lack the information needed to recognize that a change violates a security requirement.
Cable reports a roughly 20–40% vulnerability rate for even the best models on the code-writing benchmark BaxBench, from researchers at ETH Zurich and UC Berkeley. Interpreting that range requires care: BaxBench evaluates backend tasks, distinguishes insecurity among functionally correct solutions and varies security prompting. The exact historical model rows and conditions behind Cable’s range are not established here, so it should not be read as the probability that arbitrary AI-generated code is vulnerable.
Two mechanisms help explain the continuing failures. First, models learn from human-written code, which already contains insecure patterns. Second, Cable’s customer observations increasingly involve contextual issues rather than obvious one-line mistakes. An authorization check may look reasonable locally while violating a company’s business rules. Knowing the programming language does not tell the model which user may access which resource, how proprietary workflows work or what the organization’s threat model assumes. Those requirements have to reach the development process somehow.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Put security into the work before the pull request
As development becomes more autonomous, the question is how to make AI capable of producing Secure by Design software. The progression runs from autocomplete, through synchronous agents in Cursor or Claude Code, to agents working for an hour or several hours and producing large changes. Each step increases the amount of work waiting for review.
Cable forecasts that, within six to 12 months of the talk, most shipped code will be reviewed by AI rather than humans. This is Corridor’s expectation, driven by code review becoming a bottleneck as code generation accelerates. It is a prediction about how teams will respond to review demand, not a reported adoption result.
Corridor’s stated approach has two connected parts: prevent vulnerabilities before the pull request, and give security teams visibility into how AI coding tools are used. Cable expects pressure for faster development to prevail, so he frames the security decision around enabling agents with guardrails rather than withholding access.
Those guardrails become more consequential when an agent’s changes can be reviewed by another agent and merged with less human oversight. Security teams need enough assurance about that process to permit greater autonomy. Prevention and visibility therefore support the same goal: making faster development acceptable without requiring a person to inspect every action as it happens.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Defender access to powerful models
Cable then turns from development practices to model access. He describes export controls affecting Mythos and Fable, and says he participated in an Alex Stamos-led letter urging the White House to lift them. His position is that the defensive benefits outweigh the misuse risks. These are dual-use capabilities: the same model can help secure a system or exploit it. Cable credits Anthropic with safeguards around Fable intended to favor defenders.
His urgency comes from the narrowing gap between closed-weight and open-weight capabilities. In the distillation attacks he describes, a provider trains on outputs from a closed-weight model to improve another model. Cable argues that this shortens the interval before comparable capabilities become more widely available. Because adversaries already have powerful models, he sees rapid access for defenders as a necessary part of the response.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
New code, shared foundations and independent customization
Cable closes with recommendations from his June 4, 2026 congressional testimony, which he describes as occurring a couple of weeks before the talk. The first is to prevent vulnerabilities in new code. For companies and government alike, accelerating development should include strengthening the process that keeps insecure changes out.
The second is to harden the open-source foundation. Public source gives adversaries an accessible proving ground: they can point a capable model at a library and search for previously undiscovered flaws. Cable assigns responsibility to both government and private companies to help secure these shared dependencies. He again emphasizes systemic work, including rewrites that reduce whole classes of future vulnerabilities, alongside individual discoveries and patches.
The third is to support an ecosystem of American-made frontier open-weight models. Cable presents access to weights as important to competitiveness and to companies that want to customize models independently. His fine-tuning example needs a narrower technical distinction: open weights permit direct control over customization, but fine-tuning is not exclusive to them. OpenAI’s historical GPT-4o fine-tuning announcement documents a provider-hosted closed-weight option; that offering later received a wind-down notice. The durable distinction is independent access to weights versus customization controlled by a provider.
Cable situates these recommendations amid rapidly changing model releases and export-control actions. His written testimony already discusses Mythos, so the hearing should not be understood as preceding every Mythos release. The closing engineering priority is less dependent on that chronology: use increasingly capable models to strengthen fundamental controls and make systems more resilient, including against vulnerabilities that future models will discover.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
From the talk
Benchmark of 392 backend implementation tasks evaluated with functional tests and security exploits, with links to code and datasets.
Research paper describing BaxBench's evaluation of backend correctness and security.
Developer survey results covering AI adoption, usage and attitudes.
June 4, 2026 written testimony on AI coding security, open-source hardening and American open-weight models; Cable's statement begins on PDF page 18.
Further reading
Official explanation of unsafe operations and the responsibilities needed to preserve Rust's safety guarantees.
Full transcript
Loading transcript…