Anthropic's Fable 5.2 tops AI index as Claude Code leaks secrets on GitHub

Anthropic releases Fable 5.2 and Mythos 5.2, topping AI index ahead of IPO

Anthropic launched Claude Fable 5.2 and Claude Mythos 5.2, sharing a model foundation but differing in access. Fable blocks risky cybersecurity and biology tasks and is open to all users; Mythos lifts some restrictions and is limited to organizations verified for Cybersecurity or Life Sciences. Pricing is $8 per million input tokens and $40 per million output, down from $10 and $50. On Artificial Analysis Index v5 at the highest compute tier, Fable 5.2 ranks first with 65 points, five ahead of Gemini 4 Carbon and six ahead of Opus 5.5. Output is up to 30 percent faster, though per-task costs average $26.50, about 2.6 times Opus 5.5.

Claude Fable 5.2 and Mythos 5.2 cement Anthropic’s AI lead just before its IPO →

Topics: AnthropicClaude

Apiiro finds over 100,000 Claude Code settings files exposed on GitHub

Security researchers at Apiiro found thousands of secrets, including database passwords, API tokens and root logins, in Claude Code settings files in public GitHub repositories. The leak stems from how Claude Code stores approval for automated shell commands: when a user selects Yes, and don’t ask again, the rule is saved in the local .claude/settings.local.json file, often with the command written as is, including embedded sensitive data. Anthropic’s documentation says the file is added to the developer’s global .gitignore, but that setting is not reflected at repo level, so the file can be uploaded accidentally. Apiiro found over 100,000 such files on public GitHub, of which about 1,100 from 63 organizations contained secrets.

Claude Code is accidentally exposing secrets all over GitHub →

Topics: Claude Code

Meta launches Muse Ember, a small model for routine agent tasks

Meta introduced Muse Ember, its smallest and most efficient AI model to date, built for routine agent work: calling tools, small coding tasks, routing requests, extracting data and processing long documents. Meta says it delivers more intelligence per dollar at every reasoning level than Claude Haiku 5.5 or GPT-6 Luna. It costs about $0.14 per task in the Artificial Analysis test, versus $2.39 for the larger Muse Spark 1.3 at a nearly identical score of 36 versus 37. Haiku still outperforms it on real coding tasks in FrontierCode 2.0-AA, scoring 34.9 percent to 27.2 percent. Muse Ember handles text, images, audio and video, processes up to one million tokens and is available through the Meta Model API and Muse Code.

Meta launches Muse Ember, a small AI model built for the repetitive work agents actually do →

Topics: Meta Muse

Google adds TPU 8tx and 8i Max to its eighth-generation lineup

Google expanded its eighth-generation TPU portfolio with TPU 8tx and TPU 8i Max, the first time it has introduced four TPU systems in a single year. They join TPU 8t for large-scale training and TPU 8i for real-time reasoning, which offer up to 2.7 times better training performance and 80 percent better inference price performance versus the previous generation. TPUs now serve nearly 100 customers. TPU 8tx targets secure and confidential inference for regulated enterprises and governments, with up to 2x lower latency than TPU 8t. TPU 8i Max pairs two low-latency chips on one package, doubling compute, SRAM and HBM bandwidth. Google also open-sourced TorchTPU and reported state-of-the-art price-performance on InferenceX and AgentX benchmarks.

Our newest TPUs: 8tx and 8i Max, built for the agentic era →

Topics: GoogleAI chips

OpenAI’s Codex desktop app forwards local SSH agent to remote machines

OpenAI’s Codex desktop app forwards the local SSH agent to remote machines, letting an agent running remotely authenticate using the user’s local SSH credentials. As a result, the remote agent can, for example, push to the user’s private repositories on GitHub. The behavior extends the agent’s reach beyond the local machine and into the user’s authenticated sessions, raising questions about how much access desktop AI tools should grant to remote workloads.

ChatGPT Desktop forwards your SSH agent to remote machines →

Topics: OpenAI Codex