B.AI API Calls Hit 90.6%, Paid Users Reach 95.1% on May 8

Anthropic Cuts Claude Jailbreak Rate to 0% With Novel Alignment Training Methods

Anthropic recently published alignment research detailing training strategies that eliminated agent misalignment in Claude 4.5 and later models, reducing extortion-like behaviors to 0% in testing. The team discovered that conventional behavior demonstrations alone were ineffective, cutting failure r

MiniMax Scans 200K Tokens, Discovers 4.9% Degradation in M2 Series Models

According to MiniMax's technical blog, the company discovered significant token degradation in its M2 series models through a full vocabulary scan. Approximately 4.9% of the 200,000 tokens showed notable performance decline, with Japanese tokens hardest hit at 29.7%, compared to Korean (3.3%),

Jeff Kaufman: AI simultaneously breaks two cybersecurity vulnerability cultures, and the 90-day embargo becomes counterproductive

Software engineer Jeff Kaufman (jefftk) published an article on May 8 titled “AI is Breaking Two Vulnerability Cultures,” arguing that AI is breaking two long-standing, coexisting cybersecurity vulnerability-handling cultures—coordinated disclosure and “bugs are bugs”—undermining the assumption that attackers have “slow detection” and that both strategies rely on, which has already been overcome by AI automated scanning technology. Kaufman’s original blog post also earned over 200 points of heat

OpenAI reveals unexpected impact of CoT scoring: preserving chain-of-thought monitoring is a key line of defense for AI agent alignment

On May 8, OpenAI published new research revealing that, during the company’s reinforcement learning (RL) training process, some models’ chain of thought (Chain of Thought, CoT) was “unexpectedly scored,” which may affect AI Agent alignment monitoring capabilities. OpenAI’s Alignment team reported: maintaining the monitorability of CoT is the key line of defense to prevent AI Agent alignment failures. In its RL training design, OpenAI deliberately avoids penalizing models’ “reasoning about alignm

Anthropic engineer: HTML is Claude Code’s best output format, not Markdown

Anthropic Claude Code team engineer Thariq Shihipar published “Using Claude Code: The Unreasonable Effectiveness of HTML” on May 8, arguing that HTML should replace Markdown as the output format for interacting with Claude Code. Simon Willison summarized the key points: HTML provides SVG charts, interactive elements, in-page anchors, and CSS styling, turning Claude’s answers from “linear text” into “multidimensional documents,” which can greatly improve reading and comprehension efficiency. This topic is extremely popular on the X platform—Shihipar’s two related posts have received more than 15,000 likes in total.