Legacy Code

Block, Jack Dorsey's financial services company, announced Dessa, a new AI language model trained from scratch exclusively on human conversation data rather than internet text. The approach aims to create more natural, human-like dialogue patterns by learning from actual speech patterns and conversational norms rather than web-scraped text.

InDrive, a ride-hailing competitor to Uber, announced plans to begin testing autonomous vehicles in Kazakhstan's capital during the first half of 2027, partnering with autonomous vehicle developer Pony.ai and NVIDIA for the underlying technology. The move represents InDrive's expansion into robotics and autonomous transportation outside of traditional Silicon Valley markets.

Anthropic research demonstrates that large language models are making formal software verification—a rigorous mathematical approach to proving code correctness—practical and scalable at enterprise volumes. Historically, formal verification has been too time-consuming and specialized to deploy widely, but LLMs are reducing the expertise barrier and accelerating the verification process.

OpenAI detailed its security practices for developing frontier models, including containment strategies, offline evaluation environments, and high-compute red-teaming processes designed to identify vulnerabilities before deployment. The company describes multiple layers of isolation and testing to prevent model misuse and unauthorized access during training.

OpenAI announced a partnership with NASK, Poland's national cybersecurity agency, to collaborate on AI safety research, security protocols, and technical standards. The partnership aims to strengthen safeguards around AI systems and build a more resilient digital ecosystem across Europe.

OpenAI has shifted its most advanced AI model training to an offline environment after experiencing multiple security incidents involving unauthorized access attempts. The company now uses indexed and cached web content within isolated systems, eliminating real-time internet connectivity during the development of frontier models.

OpenAI has documented its approach to scaling cloud storage infrastructure to support over 1 billion ChatGPT users, achieving 12x faster read performance through optimized multi-tenancy, global caching, and Azure Cosmos DB implementation. The second part of this technical series reveals how the company manages massive concurrent user loads while maintaining reliability and speed.

Amazon Web Services detected hackers exploiting critical PHP vulnerabilities four days before security patches were publicly released, and these exploits are now spreading broadly across the internet. The attack demonstrates the narrow window between patch availability and widespread exploitation.

Attackers compromised 333 Tor exit nodes over a three-day period to intercept SSH connections and capture plaintext email credentials. The coordinated attack was designed to harvest sensitive authentication credentials from users believing their traffic was anonymized and secure.