Cerebras files for a second IPO attempt at a $23 billion valuation

Cerebras filed for its second IPO attempt targeting a May Nasdaq listing at a $23 billion valuation, backed by major deals with AWS and OpenAI. Meanwhile, three critical command injection vulnerabilities in Claude Code’s CLI surfaced from last month’s source code leak, threatening credential exfiltration in CI/CD pipelines, and independent analysis of Claude Opus 4.7’s system prompt revealed Anthropic’s evolving approach to model behavioral constraints.

Funding & Business #

Cerebras Files for IPO, Targeting May Nasdaq Listing #

TechCrunch

AI chip startup Cerebras filed a registration statement for a proposed IPO on Nasdaq under ticker “CBRS,” its second attempt after withdrawing a 2024 filing due to regulatory scrutiny of UAE-based investor G42. The company reported $510 million in 2025 revenue with $237.8 million in net income and is valued at $23 billion following its February Series H. Recent deals with AWS to deploy Cerebras chips in Amazon data centers and a reported $10 billion-plus agreement with OpenAI position the company as a credible alternative to NVIDIA for AI inference workloads.

Developer Tools #

Claude Code Source Leak Reveals Critical Command Injection Vulnerabilities #

BeyondMachines / Lobsters

Three command injection vulnerabilities (collectively CVE-2026-35022, CVSS 9.8) were discovered in Claude Code’s CLI following last month’s accidental sourcemap leak of 512,000 lines of TypeScript. The flaws affect the command lookup utility, editor invocation subsystem, and authentication helper, enabling arbitrary code execution and credential exfiltration of AWS, GCP, and API keys. The authentication helper vulnerability is particularly dangerous in CI/CD environments because it bypasses the agent’s security sandbox. Users should update immediately and set ANTHROPIC_API_KEY directly as an environment variable rather than using authentication helpers.

Changes in the System Prompt Between Claude Opus 4.6 and 4.7 #

Simon Willison

Simon Willison analyzes the system prompt differences between Claude Opus 4.6 and 4.7, taking advantage of Anthropic’s unique practice of publishing system prompts for their user-facing chat systems. The archive now dates back to Claude 3 in July 2024, providing a rare longitudinal view of how a major AI lab iterates on model instruction and behavioral constraints across releases.

Claude System Prompts as a Git Timeline #

Simon Willison

Building on Anthropic’s published system prompt archive, Willison used Claude Code to split the Markdown page into separate files per model and date, then committed them chronologically to produce a git-diffable timeline of prompt evolution. The project makes it straightforward to track exactly what changed between any two system prompt versions – a practical research tool for anyone studying how system prompts shape model behavior.

Other #

Tesla Expands Robotaxi Service to Dallas and Houston #

TechCrunch

Tesla launched unsupervised robotaxi service in Dallas and Houston, expanding beyond its initial Austin deployment to three Texas cities. The expansion is notable for skipping the supervised phase – Tesla went directly to fully driverless operation in both new markets, unlike the gradual rollout in Austin. However, fleet size appears minimal (crowdsourced data suggests one vehicle registered per city versus 46 in Austin), and Austin’s robotaxis have been involved in 14 crashes since launch. The aggressive geographic expansion with minimal fleet suggests Tesla is prioritizing regulatory precedent over service volume.

Anthropic’s Claude Mythos Launch Is Built on Misinformation #

AI Made Simple / Lobsters

A detailed critique argues that Anthropic’s Mythos launch overstated the model’s autonomous capabilities with specific methodological concerns: the “thousands of severe zero-days” claim extrapolated from only 198 manually reviewed reports; independent testing showed multiple smaller models could replicate the highlighted FreeBSD vulnerability detection; the published exploit transcript reveals substantial human guidance rather than autonomous discovery; and the Firefox exploit was tested without the browser sandbox. Five of eleven launch partners are also Anthropic investors, and JPMorgan serves as both launch partner and IPO underwriter. Whether or not every criticism lands, the piece raises legitimate methodological questions that prior coverage of Mythos’s opaque CVE count did not.

Threads to Watch #

AI tool security debt is compounding. The Claude Code CLI injection flaws join last week’s MCP design vulnerability, Git identity spoofing, and Chrome exploit stories to form a pattern: as AI-powered developer tools embed deeper into workflows, the attack surface grows faster than security tooling can follow. The source code leak that enabled these findings was itself an operational security failure, suggesting that the entire AI tooling supply chain needs hardening.

Scrutiny of frontier model claims is intensifying. The Mythos misinformation critique, combined with ongoing questions about Project Glasswing’s opaque CVE count, signals a shift from accepting lab announcements at face value toward demanding reproducible evidence. This is healthy for the field, but it creates a tension: labs that invest in safety-motivated capability restrictions face criticism both for overstating and for understating what their models can do.