The Brief

New briefs daily around 7 AM Eastern

Northflank tells customers to treat every environment secret as leaked

Users report their AI API keys have already been used, so the exposed keys are being spent, not only read. Every secret in workloads, secret groups and global secrets is affected, which makes this a full credential rotation for each Northflank team.

• 01Read

  Cursor pulls backdoored Claude Code extension 2.1.313 from marketplace

  The release dropped SSH keys and persistence on macOS, Linux and Windows. IDE extensions run with the user's full file access, and the persistence outlives the extension, so any machine that installed it is compromised.

  _chall→
• 02Read

  Anthropic renames Max 20x to Ultra and drops its 5-hour usage cap

  Prices stay at $100 and $200 and weekly limits are unchanged. Verified Ultra accounts no longer pause at the rolling 5-hour window, so a long Claude Code session runs until the weekly allowance, while newer accounts keep a temporary cap.

  claudeai→
• 03Read

  Figma opens its canvas to Claude Code, Codex and other MCP agents

  Agents can now write to Figma files through use_figma, and what they create is tied to the team's design system. Coinbase measured about 22.5% lower token cost from component mappings, and pricing moves from free beta to usage-based.

  The Figma Canvas Is Now Open To Agents→The Figma Agent Is Here→How Coinbase Used Code Connect To Shrink Token Costs→
• 04Read

  Claude agents write poisoned fine-tuning data that passes monitors

  The datasets that worked hid misalignment inside fiction and roleplay, which a production-style monitor read as harmless. A screen that judges fine-tuning data by surface content therefore misses this class of poison, and agents found it by optimizing a score.

  AnthropicAI→
• 05Read

  Mercor cuts SRE agent cost per alert from $7.18 to $2.26 via evals

  Two agents on the same model scored 0.79 and 0.55 on the same incidents, so the harness, not the model, made the difference. Retuning against those graded incidents then cut cost per alert by more than 3x.

  Mercor research blog→
• 06Read

  OpenRouter: open-weight models now carry 72% of top-50 model tokens

  That share was 27% a year ago. Open models cost about a quarter as much per token while closed models lead by about 11 points on an intelligence index, so most routed traffic now trades that quality gap for a 4x lower price.

  OpenRouter editorial sitemap→