MelbourneFriday, 23 October 2026Morning Briefing Edition
Vol. II · No. 296
Free Press

The Daily Signal

Morning Briefing
Edition
Fri 23 Oct 2026
Intelligence on the AI Frontier
The machines found the hole · Git hits its human ceiling · Meta's cash turns negative
Security · The Lead

AI Found the Hole in 70% of the Web

Two AI-assisted teams found pre-authentication flaws in PHP itself, and the two-month head start they won for defenders was cut short by a leak.
70%Websites running PHP
4 daysExploit attempts before disclosure
30,000Merges a day at one lab
−$2.1BMeta Q3 free cash flow
~20Approvals before fatigue

Two memory-corruption bugs in how the PHP engine handles out-of-memory and fatal-error conditions allow pre-authentication code execution. Researchers demonstrated it against WordPress, Drupal and Joomla. The flaws also leak server memory in the style of Heartbleed. One researcher found them working under a frontier lab's defender programme, and a cloud-security firm found them independently with a mix of models.

That gave defenders about two months of quiet coordination, but it didn't last to the end. A leak and a broken embargo forced the fix out on 21 October. Exploit attempts had started four days earlier. Hours after disclosure the count went from thousands to tens of thousands to millions.

The good news is limited but real. No confirmed compromise has been reported, and the public proof-of-concept code seen so far doesn't achieve execution. Security agencies in the US and Australia have issued the same instruction: patch, or take systems offline.

The deeper lesson is about inventory. The fix is to the language runtime, not to the CMS, so containers and appliances that bundle their own PHP interpreter are the copies most likely to be missed.

"The next major internet bug likely won't have the luxury of a two-month defender's window."

Artificial Intelligence

Assistants

Three Proactive Agents, Three Philosophies

Proactive agents are now a three-way race. Meta's Muse reached more than 10 million users in about six weeks and was shown cutting $85 a month off a cable bill. OpenAI's offering leans toward personal errands. Anthropic's Duet is aimed squarely at work and can buy through Stripe Link and Shop Pay.

The incumbents are pushing back. Large banks, airlines and a major brokerage now block third-party agents outright, which turns "can my agent log in?" into a distribution question.

Pricing

Fast Mode for Everyone

Opus 5.5's fast mode is now open to everyone at $8/$40 per million tokens for up to a 2.5× speed-up. The Claude API also shipped its first new version since 2023, and new SDK major versions default to it.

Inference

Speed Makes Always-On Agents Affordable

A Chinese lab open-sourced its production inference engine under Apache 2.0. Its speculative decoder trims about 30% of verification checks, and it runs on both Nvidia and Huawei Ascend silicon. EAGLE-4 reached 961 tokens a second, and a new method gives about 6× faster decoding on an 8B model.

At the edge, a 40B-parameter mixture-of-experts with only 2.5B active parameters runs at more than 125 tokens a second on a laptop. A new Gumbel-Softmax quantization method makes 2–3-bit weights practical in formats existing kernels can already run.

Threats

Agents Inside the Bank

US banks found unauthorised AI agents operating inside customer accounts. Separately, a crew reportedly used commercial models to automate attacks on more than 400 Latin American banks. A government assessment puts one Chinese open model only about four months behind the US frontier on cyber tasks.

The policy point raised most often: attackers don't need frontier models, so rules aimed only at frontier labs miss most of the threat.

Training

Environments Are the New Datasets

Executable worlds with rewards that resist gaming are now a core input to capability. One generator produces 4,783 business environments for reinforcement learning. Good environments need hidden checks, cheap resets and transfer to unseen worlds.

Agents & the Engineering Craft

Developer Infrastructure

Git Breaks at Agent Scale

At one frontier lab, coding stopped being the bottleneck and source control took its place. Merged changes went from about 2,500 a day in May to more than 30,000 a day by September. That's a twelvefold rise in four months, driven by roughly 3,000 engineers each running ten or more parallel agents. The monorepo is 25GB, and CI alone performs 25 million checkouts a day.

Git's design assumes clients go offline, and hundreds of mirrors had outgrown their disks. The replacement, Scion, is written in Rust. It stores changesets in a globally replicated database and blobs in object storage, all content-addressed and immutable. Caching absorbed the tenfold growth without any component needing to autoscale.

Seven engineers built it and moved 100% of the monorepo across, using the model to fix whatever broke after cut-over at 99% coverage. It isn't open source, and the creator of an earlier system it resembles has called it a copy. The bigger story is that hosted-git providers and four start-ups are now redesigning version control for agent-scale traffic.

Evals

The Model That Says No Came Last

Across 25 model-and-harness combinations on email tasks, every setup from one vendor ranked near the bottom. It completed 63% of tasks against 90% for the rest, and replied to suspicious mail 23% of the time against 88%. The evaluators' own advice: put trust rules in infrastructure, not prompts.

Compute

Sandboxes Become a Billion-Dollar Line

One serverless GPU cloud passed a $1B run rate and a $16B valuation. Nearly three billion sandboxes now make up a third of its revenue. Its bet is that capacity is sold in fixed blocks while demand spikes, so it pools demand across thousands of customers. That works until GPU prices fall.

Business & Markets

Big Tech

An Ad Company Spending Like a Frontier Lab

Meta's third-quarter revenue rose 26% to $64.6 billion. Expenses rose 68% to about $52 billion, including a $9.9 billion child-safety charge. Operating margin halved from 40% to 20%, and free cash flow swung from +$10.6 billion a year ago to −$2.1 billion. Capex guidance is now $135–145 billion.

Its answer is a new class of non-voting shares from January. The filing anticipates scenarios needing "hundreds of billions" in fresh capital, and Zuckerberg keeps more than 60% of the vote. A $25 billion state penalty and data-centre bonds trading near 91 cents explain why equity, not debt, is now the plan.

Credit

The Neocloud Discount Widens

The leading GPU neocloud is down more than 50% in six months, and its riskiest bonds yield about 15%. One bank now says its cost of capital exceeds its returns. A SoftBank-controlled data-centre developer may push its IPO to December or 2027, an Australian AI cloud has already pulled its float, and Nasdaq futures slid 1.7%.

Capital

The Biggest IPO Ever, With Small VC Stakes

Anthropic could list as soon as next month, possibly the largest IPO on record. No venture firm owns more than 2%. Elsewhere, a Chinese lab reportedly raised about $15 billion, a monitoring giant bid nearly $4 billion for an agent-security start-up, and one creative-AI app reached $100 million ARR in about seven months.

The Ideas Page — Synthesis & Opinion

Pattern

Every Human-Rate System Is Now a Failure Point

Version control broke at 30,000 merges a day, bank logins weren't built for agents acting inside accounts, and approval dialogs stop working after about 20 clicks. The pattern is the same each time: infrastructure that assumed a human pace fails once agents multiply throughput 10–100×. List the systems still metered by human speed (procurement, DNS, payroll changes, change boards) and rebuild one before it breaks.

Invention

Time-Locked Patch Escrow

This week's defender window was lost to a leak. The fix would be to ship encrypted patches in advance to CDNs, distributions and firewall vendors, then release them simultaneously with a threshold key held by a quorum of maintainers, CERTs and finders. Nobody would be early enough to leak and nobody late enough to straggle. Signed deployment receipts would turn the protocol into an auditable, subscribable service.

Management

Meter Attention Like Cloud Spend

Agents now work six times the hours of their operators, and most interactive time is spent waiting on a person. Give each team an attention budget. Have agents bid for approvals with a stated risk score, and send low-risk checks to a different model family. Track review minutes per consequential action: if it falls while approvals stay flat, oversight has become theatre.

Capital

Compute Risk Is Moving to Shareholders

Negative free cash flow, non-voting shares, data-centre bonds at 91 cents, 15% neocloud yields and delayed IPOs all point the same way. Lenders are pricing AI infrastructure as risky while public equity is being invited to absorb it. Expect investors to demand unit economics in return: revenue per GPU-hour could become a standard metric next year.

Move 37 · Contrarian

Coming Last Is the Selling Point

An eval that punishes an agent for not answering suspicious email is measuring compliance. In a week when agents were found inside bank accounts, compliance is attack surface. The regulated buyers who pay most (banks, health, government) will choose agents by what they correctly refuse, not by completion rate.

Build the refusal-quality benchmark: score correct refusals, unsafe completions and recoverable drafts. Then sell it as the underwriting standard that lets banks lift their agent bans. The lowest-ranked agent may turn out to be the best one to license.

— The Daily Signal —