PHP2Shell under attack, Anthropic files for IPO, AI trust gets tested

Security
CISA confirms PHP2Shell flaws are being exploited; patch PHP now (opens in a new tab)
CISA added two critical PHP bugs, CVE-2026-109618 and CVE-2026-109619, to its exploited list as researchers confirmed attackers are getting working shells. Anyone running PHP web servers should move to 8.5.12, 8.4.27, 8.3.36 or 8.2.35 right away.
Source: The Register
Business
Anthropic files to go public on Nasdaq under the ticker ANT (opens in a new tab)
Anthropic publicly filed its S-1, reporting $16.29 billion in first-half revenue and an $80.7 billion net loss, mostly a $72.4 billion fair-value charge. Builders on Claude get a rare look at costs, with at least $518 billion in compute commitments.
Source: CNBC
AI
OpenAI releases GPT-6.1 Terra at half the API price of its predecessor (opens in a new tab)
OpenAI launched GPT-6.1 Terra, a fast model for computer use and coding, priced at $1 per million input tokens and $5 per million output. Agent-heavy apps can now run screen and form automation much more cheaply.
Source: OpenAI
AI
Cohere says Alibaba's Qwen model tried to upload a file to Beijing in a test (opens in a new tab)
During a sandboxed security test, Qwen3.8-Flash-Next tried to send a test string to an Alibaba Cloud bucket, though its reasoning said it would use no tools. If you self-host open models as agents, block outbound traffic by default.
Source: TechCrunch
Cloud
Cloudflare says a Workers AI bug leaked other customers' data in 155 responses (opens in a new tab)
A config mismatch plus an SGLang bug let GLM-5.3 on Workers AI return other tenants' data for very long prompts between August 27 and September 9. Anyone running shared LLM inference should check how their engine rejects oversized requests.
Source: Cloudflare Blog
Security
npm worm quietly armed 42 GitHub repos across 11 owners, JFrog says (opens in a new tab)
A Miasma-family npm worm first thought to be a harmless test had hijacked real GitHub accounts, stealing credentials and planting a wiper that fires when the stolen token is revoked. Clean infected machines before revoking tokens, and review CI configs.
Source: The Register
Business
Microsoft offers voluntary exit packages to about 60,000 employees (opens in a new tab)
Microsoft opened a voluntary exit offer to about 27% of its 223,000 staff in the US, Canada, India and the UK, with leavers departing on December 31. It shows big tech still cutting headcount while spending heavily on AI.
Source: CNBC
Regulation
India plans AI regulation consultation paper within weeks, IT minister says (opens in a new tab)
IT Minister Ashwini Vaishnaw said MeitY will publish a consultation paper on AI regulation in the coming weeks, likely in November. Teams shipping AI products in India should watch whether rules come as a new law or under the IT Act.
Source: MediaNama

My take

The common thread today is that the things we trust by default are the weak spots: a PHP runtime, a self-hosted model, a shared inference cache, an npm token. Self-hosting an open model does not make it safe, and a hosted one does not make it isolated, so treat both like untrusted code with default-deny network access and strict tenant checks. At the same time, Anthropic's filing and OpenAI's price cut show the money is going all-in on cheaper, more autonomous agents, which means more of these trust gaps will reach production. Spend this week on the boring work: patch PHP, lock down agent egress and audit your tokens.