Meta's guardrails cracked in 10 minutes

Financial Times

Safety guardrails on Meta and Google's open-source AI models can be stripped in under 10 minutes with free tools, exposing significant vulnerabilities in how these models are protected against misuse.

Wired

Anthropic's Claude Code and open-source OpenClaw have created 'Claudeholics'—developers reporting 90x productivity gains but burning through AI budgets at unsustainable rates, with per-developer token costs quadrupling to $2,000/month while ROI remains invisible.

TechCrunch

ClickUp laid off 22% of staff while deploying 3,000 AI agents to create a 3:1 AI-to-human workforce ratio, promising million-dollar salaries to employees who leverage AI effectively. The move reflects a broader trend where 80% of companies using autonomous tech have reduced headcount, though productivity gains remain unproven for many.

Ars Technica

A trivial-to-exploit vulnerability in Starlette framework exposes millions of AI agents and servers to authentication bypass, threatening credentials stored in MCP servers and affecting major tools like FastAPI, vLLM, and LiteLLM. The flaw was patched Friday but poses critical risk to the Python AI infrastructure ecosystem.

TechCrunch

AI gateway OpenRouter raised $113M at a $1.3B valuation, more than doubling in value within a year as it processes 100 trillion tokens monthly across 400+ models. The startup's success indicates enterprises are embracing a multi-model strategy rather than standardizing on single AI providers.