10 Comments
User's avatar
Jojo's avatar

Whew. I'm so glad I am no longer working and don't have to sit through all the meetings after meetings after meetings discussing critical subjects like this!

olgreenie's avatar

Think you accidentally neglected to name a certain load-bearing co-author on this one, I work with him too

JG's avatar

If you are going to publish a long-form article for public consumption, please include in an upfront disclaimer if you used gen AI to write most of it. Getting through the first few paragraphs it was clear this was almost entirely AI-generated. The prose is tedious, verbose, and just difficult to read; all the tell-tale signs of a Claude-written piece are here. It's great research diluted by generic writing. I expect better from an institution like a16z.

Latent Dynamics's avatar

Trying to patch your way out of machine-speed vulnerabilities is a mathematical dead end. 🛑

Look at the raw physics of the current crisis. When an unreleased frontier model synthesizes a remote, zero-login Linux kernel exploit in under 60 minutes, your traditional human-gated patch cycle, which averages 43 to 252 days according to breach audits, has already failed. You aren't operating in a security race. You're suffering from a catastrophic clock-rate mismatch. 📉

Generative synthesis operates at hardware speeds (τ_exploit ≈ 10³ ms), while enterprise change boards operate on calendar schedules (τ_patch ≈ 10⁹ ms). Attempting to close that spectral gap by rushing software deployments just swaps external intrusion for self-inflicted CrowdStrike-style outages.

The structural flaw sits in where we place the verification plane. Source code isn't the live source of truth anymore. Production is. The moment an exploit payload hits an exposed port, asking a human engineer to review a pull request is pure ceremony. ⚙️

True immune state invariance requires enforcing an out-of-band physical bus interposer. We must compile live threat advisories directly into microsecond hardware access-control gates before a single byte touches memory. 🛡️

Instead of rewriting legacy application layers, defense must migrate to deterministic execution envelopes:

1. In-Process Runtime Hooks: Lightweight enclaves (using eBPF or Landlock) intercept untrusted input at core parsers, evaluating exploitation patterns in microsecond windows (τ_gate < 15 µs) without host CPU stalls.

2. Per-Request Principal Attribution: Autonomous agents on developer laptops break standard session tokens. Trust must be evaluated per request, requiring short-lived cryptographic credentials signed at the endpoint driver layer.

3. Immutable Out-of-Band Restores: Backup posture isn't an operational assumption. If recovery isn't tested via automated clean-room restores, it's just zero-value marketing. Sophos numbers prove that un-isolated backups cause median recovery costs to jump eightfold from $375k to $3M.

Stop trusting probabilistic software promises to guard open ports. Isolate the execution plane, gate the requests at the network edge, and commit state changes strictly through deterministic hardware boundaries.

(╯°□°)╯︵ ┻━┻

Steve Garrison's avatar

Who stands between the attacker and the unpatched pump for ten years? Shields buy time, but somebody still has to notice the chain. (I'm at Stellar Cyber.)

Hideaki Takahashi's avatar

Thank you so much for the detailed article!!! Formal verification is definitely a great tool to eliminate entire bug classes, and I think it's getting dramatically faster. SMT-based formal verification has been automated from the beginning, though it’s still slow due to the inefficiencies of bit-vectors. Meanwhile, the other approach, interactive theorem proving, is improving fast. Lean4-style theorem proving was not popular before since we human needed to write the proof by hand. However, in 2026, we can just ask AI to write those proofs, and now an "interactive theorem prover" is mostly automated.

Ollie Darby's avatar

the industrial control systems bit is the one I'd underline, and it made me think of Onodrim who just raised $300m to buy up European defence and industrial manufacturers, because if those factories are running controllers that only get patched once a year then the cyber side might be one of the first things a buyer like that has to sort out

Abhishek Pan's avatar

Anthropic's unreleased model winning a Linux race that Red Hat rated moderate in under an hour is the new clock. Severity that assumes a human needs days to win a timing race goes stale the moment a model tries it, so the compensating control has to land before the exploit write-up does.

Nan Deng's avatar

The number that sticks is the 43-day median full remediation against exploits that can be assembled in hours. The PHP week is the concrete case: edge providers could block while the underlying installs stayed open, and 2.6–4.4 million installs outside that umbrella stayed exposed. I would treat "can someone stand between the attacker and the bug" as a first-class control, not a temporary patch. Otherwise every unfixed package is just waiting for the next model that wins the race condition in under an hour.

AI Delivery Brief's avatar

The gap in this piece is the one to sit with: exploits take hours, fixes take weeks, and the answer is to ship fixes at machine speed too. But a bad fix shipped at machine speed is self-inflicted harm. So the checking has to travel at the same speed as the fix. And a growing share of breaches now start with a vendor — whose risk still sits with the buyer. https://aideliverybrief.substack.com/p/mas-says-a-vendors-ai-is-still-your