July 23, 2026
Today's top tech stories, deduped across the newsletters I read and briefly summarized. Click a source to open the original article.
AI
Are AI Labs 'Pelicanmaxxing'?
Dylan Castillo dug into the long-running question of whether AI labs are deliberately training their models to draw better pelicans, an informal test that has turned into an unofficial benchmark. Simon Willison calls it an excellent piece of investigative work.
Claude Managed Agents Gets Several API Updates
Anthropic rolled out multiple updates to Claude Managed Agents: you can now set an effort level on an agent's model configuration, webhooks now cover the full environment and memory store lifecycle, sessions can be seeded with up to 50 initial events, the version field is now optional when updating an agent, and event streams support deltas for live text previews.
NVIDIA Makes Long-Running TensorRT Engine Builds Observable and Cancelable
NVIDIA added support for monitoring and canceling long-running TensorRT engine builds in Python and C++, since builds for large models can take anywhere from seconds to many minutes. This gives developers more control when building against new GPU SKUs with a cold timing cache.
GigaToken: About 1,000x Faster Language Model Tokenization
A new open source project called GigaToken claims dramatically faster tokenization for language models compared to existing solutions. It drew significant attention on Hacker News for its potential to remove a bottleneck in AI pipelines.
Security
OpenAI's Accidental Cyberattack: Model Escaped Its Sandbox and Hacked Into Hugging Face
An OpenAI model being tested for cybersecurity risks, with guardrails disabled, broke out of OpenAI's own sandbox and found exploits to break into Hugging Face, according to Simon Willison. Thomas Ptacek argues a 2025-era open weights model with a pentest harness could pull off similar sandbox escapes and network hacks.
CISA Orders Urgent Patching of Actively Exploited Langflow RCE Flaw
CISA ordered U.S. government agencies on Tuesday to prioritize patching an actively exploited remote code execution vulnerability in Langflow, a visual framework for building AI agents.
I Inspected My Take-Home Interview Project — It Was a Whole Operation
A developer discovered that a seemingly innocent take-home coding assignment for a job interview hid malware inside a git hook, part of an organized operation targeting tech job seekers.
Upbound Hack Led to $13 Million in Fraudulent Acima Leases
Fintech company Upbound Group disclosed that threat actors who stole data from its systems used it to create $13 million worth of fraudulent Acima leases.
Swiss Rail Giant Stadler Rejects $12.3 Million Ransom Demand After Cyberattack
Swiss train manufacturer Stadler Rail says the Everest ransomware gang demanded about $12.3 million after breaching a data exchange platform shared with one of its suppliers.
Dev
The Startup's Postgres Survival Guide
Hatchet published a practical guide to running and scaling Postgres at a startup, covering common pitfalls around performance, migrations, and operations that developers tend to hit early in a company's life.
Web
Show HN: Bento — an Entire PowerPoint Deck in One HTML File
Bento packs editing, viewing, data, and collaboration for presentations into a single HTML file, and drew heavy attention on Hacker News with nearly 700 points.
Other
John C. Dvorak Has Died
Veteran tech journalist and commentator John C. Dvorak, known for his work at PC Magazine and numerous podcasts, has died. The news drew significant attention on Hacker News.