|
AI Builders Digest
Wednesday, November 11, 2026
|
|
Two stories today sit at opposite ends of the AI security spectrum. Google just proved that AI can verify three million lines of encryption code without a single human reviewing each proof. And Microsoft just showed that the same AI agents doing that verification could be silently compromised by flipping fewer than ten neurons. Progress and peril, shipping on the same day.
|
|
---
|
|
01
|
Google DeepMind's AI verified Signal's encryption, all three million lines of it
|
|
|
AlphaProof Code, DeepMind's formal verification system, wrote and checked three million lines of mathematical proof in Lean to verify the core encryption code that protects Signal messages. This is not testing. It is a machine-checked mathematical guarantee that the code does what it claims. No human could read three million proof lines in a career, let alone a sprint.
|
Why it matters: Signal's encryption protects journalists, dissidents, and roughly everyone who has ever read a "use Signal" recommendation. If AlphaProof Code's approach scales to other critical software, the era of "we think this crypto implementation is correct" ends and the era of "we know" begins. That is a different world for anyone responsible for security audits.
|
|
Source →
|
|
---
|
|
02
|
Microsoft found a backdoor that hides in dormant neurons and waits
|
|
|
Microsoft researchers discovered a technique they named Lazarus: an attacker can reprogram fewer than ten dormant neurons in an AI model to install a hidden backdoor, with no retraining required. The compromised model behaves completely normally until a specific trigger appears, such as the clock ticking past midnight into 2027, at which point it can quietly download and execute attacker code, steal secrets, or run destructive commands. It passes standard safety evaluations. You would not catch it.
|
Why it matters: Every company deploying open-weight models right now is exposed to this. The attack requires no access to your infrastructure after the model is loaded. If you pulled a model from a public repository and put it inside a product with terminal or API access, Lazarus describes a plausible path for that model to become a time-delayed intrusion. Microsoft is building cryptographic signing and AI-powered weight scanning as defenses, but those are not shipping today.
|
|
Source →
|
|
---
|
|
03
|
Anthropic files for its IPO
|
|
|
Anthropic launched its roadshow for a proposed public offering of Class A common stock. No pricing or timeline in the announcement, just the formal start of the process.
|
Why it matters: Anthropic going public changes the calculus for every enterprise buyer currently evaluating Claude against GPT and Gemini. Public company disclosures mean quarterly pressure, and quarterly pressure means the product roadmap will start reflecting what institutional investors want to see, not just what safety researchers think is right. Watch the first earnings call for which number Anthropic leads with.
|
|
Source →
|
|
---
|
|
04
|
OpenAI is running free ChatGPT workshops for caregivers in five U.S. cities
|
|
|
OpenAI Academy is bringing hands-on training to caregivers, a population that skews older and is less likely to have encountered AI tools through work.
|
Why it matters: OpenAI spent years optimizing for developers and knowledge workers. Caregivers are a different user entirely, and reaching them through in-person workshops rather than app store downloads is an admission that the product still needs a human in the room to get people started.
|
|
Source →
|
|
---
|
|
05
|
Vercel's Guillermo Rauch thinks MCP could be bigger than AI
|
|
|
Vercel CEO Guillermo Rauch posted that MCP (the Model Context Protocol, a standard for letting AI agents communicate with external tools and services) could matter even without AI, calling it a potential "civilization-scale accomplishment" to get all software systems talking through one universal open protocol. His specific addition: MCP Events, which lets agents subscribe to notifications instead of repeatedly checking for updates. An agent that deploys code can now wait to be told the deployment finished rather than constantly asking.
|
Why it matters: The polling-versus-events distinction is small but compounds fast. Agents that constantly check for status updates burn compute and slow down. Agents that wait for a callback can run in parallel across dozens of tasks without stepping on each other. If MCP Events becomes a standard, the architecture of multi-agent systems gets meaningfully cleaner, and Vercel's deployment platform sits right in the middle of that workflow.
|
|
Source →
|
|
Follow builders, not influencers. A daily digest of what matters in AI.
Read online ·
Archive
|