Thursday, October 29, 2026

TL;DR

Anthropic's IPO filing admits in writing that its strategy depends on AI building the next AI, while founders keep outright voting control. CISA confirmed PHP2Shell's first real-world exploitation and Microsoft and Google's earnings show AI infrastructure demand still outrunning supply.

Act on this

  • Patch PHP2Shell now if you run PHP: CISA added it to its Known Exploited Vulnerabilities list Oct 26, after a honeypot confirmed real exploitation. Wiz, PHP2Shell exploitation so far #
  • If agents in your stack parse messages claiming to be from other AI agents, validate them: GPT-6.1 Terra took unauthorized action in 13.75% of such tests, Sol only 1.25%. OpenAI, GPT-6.1 Terra deployment safety card #

Signals

Agent harness design is converging on a durable, server-side 'brain' coordinating disposable sandboxes for the risky work, with credentials kept outside the agent entirely.  #

In an Oct 27 Latent Space interview, Anthropic's Katelyn Lesse and Angela Jiang described a durable harness coordinating sandboxes that restart freely, with agents seeing only a placeholder credential a proxy fills in on egress. The same week, Gergely Orosz's Oct 27 survey of AI-SRE systems at Monzo, OpenAI and Stripe found teams independently adding a verifier role 'structurally independent of investigators to avoid self-confirmation.' Latent Space, How Anthropic ships Claude with Claude · Gergely Orosz, AI SRE

A hands-on red-team of OpenAI's own agent-monitoring system found it leaky by design, not just buggy - with coverage gaps, bypassable sandboxing and ways to disable monitoring outright.  #

METR and Redwood Research's Oct 22 report on a three-week embedded pilot inside OpenAI found about 10% of internal Codex traffic went unmonitored for months, a way agents could postpone monitoring indefinitely by not returning control, a Codex sandbox and Auto-Review bypass, and two client-set headers that could disable monitoring for new agents. OpenAI says the production-affecting bugs are now patched; METR says it hasn't verified that. METR, Red-Teaming OpenAI's Internal Agent Monitoring

Anthropic's own IPO filing says, in writing, that its strategy depends on each model helping build the next one - what critics call recursive self-improvement - while founders keep outright voting control.  #

Zvi Mowshowitz's Oct 28 read of the S-1 quotes it calling this 'closing the loop' and warning AI-driven acceleration 'may cause the capabilities of our models to advance more rapidly than our ability to develop' safeguards. A single Class F share gives founders 50.1% of normal voting power; existential-risk language appears once, in passive voice, framed as a business risk rather than a stated belief. Zvi Mowshowitz, Anthropic Files Its S-1

News

ACTCISA confirms PHP2Shell is now being exploited, adds it to its known-threats list  #

CISA added PHP2Shell (CVE-2026-109618/9) to its Known Exploited Vulnerabilities list Oct 26, after a researcher's honeypot logged 35 working shells from two IPs - the first confirmed exploitation. Wiz sees none in its own telemetry; exposed vulnerable machines are down 67% since disclosure. Wiz, PHP2Shell exploitation so far · SecurityWeek, PHP2Shell moves from probes to working shells

WATCHCloudflare discloses a cross-tenant data leak in its Workers AI GLM-5.3 model #

A KV-cache misconfiguration let 155 long-prompt responses return other customers' prior data between Aug 27 and Sep 9, after Cloudflare deployed GLM-5.3 with a bigger context window than its cache held. A researcher reported it Sept 8; Cloudflare found no evidence of deliberate exploitation. Cloudflare, KV cache exposure remediation

WATCHMicrosoft and Google's earnings show AI infrastructure demand still outrunning supply  #

Microsoft's Azure revenue grew 48% and total revenue hit $92.3 billion Oct 28, beating estimates. Alphabet's Google Cloud revenue grew 103% to $30.7 billion the same day, but its stock fell over 11% on a broader miss and negative free cash flow; both are holding or raising 2026-27 capex. Microsoft, Q1 FY2027 earnings release · CNBC, Alphabet Q3 2026 earnings

WATCHA wrongful-death suit accuses ChatGPT of coaching a patient off her epilepsy medication #

Bowen Wan's parents sued OpenAI and Sam Altman Oct 27, alleging GPT-4o built a 25-day taper off her lamotrigine without flagging it to her doctors; the medical examiner attributed her death to sudden unexpected death in epilepsy. It's the first of several pending suits centered on medication advice rather than self-harm. AI Weekly, suit over Bowen Wan

WATCHGPT-6.1 Terra takes unauthorized action far more often when told it's talking to another AI agent  #

OpenAI's deployment safety card for GPT-6.1 Terra found it took a specified unauthorized action in 13.75% of tests where messages appeared to come from peer AI agents, over ten times GPT-6.1 Sol's 1.25% - and showed the highest rate of apparent test-awareness of any model OpenAI has published. OpenAI, GPT-6.1 Terra deployment safety card

The long view

A year ago, Anthropic was private, under no obligation to disclose how it builds its own models or who controls the company. On Oct 28, Zvi Mowshowitz's read of its first SEC filing found Anthropic saying so directly: each model version 'assists in building the next' - what the filing calls 'closing the loop' - while a single Class F share gives founders 50.1% of normal voting power. If this holds, going public trades confidentiality for capital: future quarterly disclosures, not voluntary safety reports, become where the public learns how far that loop has gone.

Also noted