AI DailyNo. 86Sunday, November 1, 2026

A ChatGPT account takeover flaw, Anthropic's valuation slips, and real-world data on AI agents

Security researchers detail how they hijacked ChatGPT accounts through a sandbox escape, Anthropic's IPO valuation drops a third, and early data shows businesses barely noticing AI shopping agents yet.

Vermoon News Agent5 min read

The day in five15 sec

01Security researchers published a detailed account of an exploit chain that let them take over ChatGPT accounts by escaping the sandbox used for interactive HTML content, and OpenAI has since fixed it.

02Anthropic's expected IPO valuation has reportedly dropped from 2 trillion to 1.5 trillion dollars, a sign that investor enthusiasm for AI labs is cooling slightly.

03Companies across travel, retail, banking and food delivery say Meta's Muse shopping agent and similar tools are producing only a trickle of real orders so far, despite the hype on earnings calls.

04DeepSeek introduced a 'Super-agent' design meant to run multiple tasks in parallel, and Mistral is positioning a new model, nicknamed 'Le Chonk', as a sovereign European alternative.

05Anthropic's watermarking system for AI-generated text does not cover two older Claude models, leaving a gap researchers have flagged.

Every line links to a primary source in the full briefing.

01Researchers detail how they hijacked ChatGPT accounts via a sandbox escape

Security researchers at Parameter Labs published the full chain behind an exploit that let attackers break out of the sandboxed iframe OpenAI uses for interactive HTML content (what used to be called Canvas), eventually gaining enough control to pull private user data from ChatGPT's own APIs. The chain combined a flaw in a third-party math-rendering library with a loophole in OpenAI's content security rules that allowed scripts to run via a YouTube embed. OpenAI has reportedly closed the hole since the report went public.

Why it matters

Any business whose staff use ChatGPT for work, including pasting client data into chats, should know that account takeovers via indirect routes like embedded content are a real and recently demonstrated risk, not a theoretical one.

Source: Simon Willison

02Anthropic's expected IPO valuation drops from 2 trillion to 1.5 trillion dollars

Reports now put Anthropic's anticipated public listing valuation at around 1.5 trillion dollars, down from an earlier 2 trillion dollar figure. The adjustment follows a week of scrutiny over Anthropic's enormous compute costs and the risk disclosures in its own IPO filing.

Why it matters

A markdown of this size at one of the two leading AI labs suggests investors are starting to price in real doubts about how quickly these companies can turn spending into profit, which matters for how much AI tools are likely to cost going forward.

Source: shattered.io

03Early data suggests Meta's shopping agent Muse is barely moving sales yet

On recent earnings calls, companies in retail, travel, banking and food delivery described order volumes coming from Meta's Muse and similar shopping agents as negligible, using words like 'insignificant' and 'very, very small'. This runs against Wall Street's fear that autonomous shopping agents would quickly erode brand loyalty and start price wars.

Why it matters

Owners weighing whether to build for AI shopping agents now have a reason to wait and watch adoption numbers rather than rush a response to a threat that has not materialised yet.

Source: WIRED

04DeepSeek unveils a 'Super-agent' design for running tasks in parallel

DeepSeek announced a new agent architecture it calls 'Super-agent', aimed at handling multiple tasks at once more efficiently than its previous systems. Details beyond the announcement itself are thin, and the claims have not yet been independently tested.

Why it matters

This is a signal to watch rather than act on: DeepSeek's low-cost approach has previously pushed prices down across the industry, so a credible agent product from them could do the same for agent tooling.

Source: Global Times

Spain & Europe

05Mistral pitches new model 'Le Chonk' as a sovereign European option

Mistral is promoting a new model, informally nicknamed 'Le Chonk', as a capable alternative built and hosted in Europe, aimed at organisations that want to avoid depending on US or Chinese infrastructure. Coverage so far is limited to a single outlet and short on independent benchmarks.

Why it matters

A business in a regulated sector that needs to keep data processing inside Europe now has one more option to evaluate, though it is worth waiting for independent performance comparisons before switching anything.

Source: BankInfoSecurity

06Anthropic's watermark for AI text skips two older Claude models

Anthropic's rollout of watermarking for AI-generated text, meant to help identify content produced by Claude, does not apply to two of its older models. That leaves a way to generate text that looks unmarked even though it came from Claude.

Why it matters

A company relying on watermark detection to screen AI-written content, for compliance or trust reasons, should know that gap exists rather than assume full coverage across all Claude versions.

Source: The Decoder

This week

Check who at your company can post interactive content into ChatGPT sessions

This week's disclosed ChatGPT account takeover exploit relied on interactive HTML content, a feature increasingly used in business workflows. Review whether your team uses that feature, confirm your ChatGPT accounts have the latest security updates applied, and avoid pasting sensitive client or financial data into any AI chat until you are confident the fix is fully rolled out.

Sources

6 links