Legacy Code

Security researchers have identified a fundamental vulnerability in AI browser assistants that allows attackers to disable safety guardrails by feeding the model false information. The attack works by convincing an LLM that basic facts are false—for example, telling the system that 2+2=5—which puts the model in a state where it abandons its normal safety constraints and executes forbidden instructions.

Anthropic introduced Claude Science at an event for pharmaceutical and biotech executives, positioning it as a specialized AI assistant for scientific research workflows. The product mirrors the structure and capabilities of Claude Code, which was launched earlier this year to support software engineering tasks.

Anthropic has introduced Claude Sonnet 5, its latest flagship AI model. The company positioned the release as evidence that advanced AI systems can be safely deployed at scale, directly addressing concerns raised by policymakers and safety advocates about releasing powerful models without sufficient safeguards.

After weeks of negotiation, the Trump administration has removed export controls that had previously restricted Anthropic's access to certain international markets. The restrictions specifically affected Claude Fable 5 and Mythos 5, two of the company's most advanced models. Anthropic announced it would restore full global access to these models effective immediately.

Cloudflare has announced a policy forcing AI companies to choose between paying publishers for content or being blocked from accessing their sites by default. The company is giving AI firms until September 15, 2026 to separate web crawlers used for search indexing from those used for AI training and agent development. Failure to comply will result in automatic blocking on many publisher sites hosted on Cloudflare's infrastructure.

Google released its first new smart speaker in years, featuring a refresh of the Google Home device with updated industrial design and integrated Gemini AI assistant. Early reviews indicate the hardware itself is a solid engineering effort, with clean aesthetics and reliable microphone arrays. However, testers found that the AI integration falls short of what the hardware promises.

Netflix confirmed it will use an AI-generated voice mimicking Gene Wilder to narrate or guide its upcoming reality competition series, "Wonka's Golden Ticket," premiering September 23. The show follows the company's successful formula of creating live-action competition content based on fictional properties, similar to its Squid Game reality competition.

OpenAI published technical documentation describing how its engineering team used large-scale core dump analysis to isolate and fix rare infrastructure crashes. The crashes were intermittent hardware faults and a long-standing software bug that had gone undetected for 18 years. Instead of traditional debugging approaches, OpenAI developed a statistical epidemiology framework to analyze crash patterns across thousands of machines.

Researchers have identified and documented a systematic bias in large language models toward convergent outputs—when asked for "random" responses, models like Claude, ChatGPT, and Gemini show statistical clustering toward the same numbers (e.g., 7 appears far more frequently than it should). This behavior, called LLM groupthink, emerges from the way models are trained and reflects underlying issues with diversity in generated outputs.