AI #191: The Hackening Begins

A critical vulnerability in PHP was announced roughly a day before a fix was available, according to Zvi Mowshowitz's weekly AI roundup, which frames the episode as possibly the worst day in cybersecurity history so far. The early disclosure stemmed from a leak rather than a coordinated release. Because PHP underpins a large share of the web, the gap between disclosure and patch sent much of the internet scrambling, and Mowshowitz expects many smaller, long-tail sites to be permanently lost. Several unrelated security incidents landed in the same week, which he reads as the floodgates opening; he says he runs a market on whether an even worse day will arrive before year end.
That is the headline item, but the post is a wide sweep across model releases, interpretability research, agent behavior, benchmarks and policy. The common thread is that capability and deployment keep outrunning the operational and institutional scaffolding around them.
A second leak involved a draft bill Anthropic had been circulating with members of Congress, described as similar to but weaker than the FRONTIER Act. Mowshowitz says Anthropic's first attempts to explain the situation made it look worse, suggesting the company did not grasp what the problem was, and that Dario Amodei's later response was better while the matter remains troubling. He plans a separate post covering the cluster of Anthropic developments, which he characterizes as ranging from excellent to terrible.
On interpretability, a new claim holds that neural networks are cleaner than assumed: individual neurons mostly carry a single interpretable meaning, and labeling them one by one is tractable. Mowshowitz says early reactions overshot, treating it as a much bigger result than it is, but he still calls it excellent news that substantially lowers his estimate of the difficulty ahead. He pairs this with a note that there have been quite a few misalignment incidents, implying the labeling work is now urgent.
On releases: Claude Duet is Anthropic's answer to proactive agents such as Muse and Dot, built on Opus 5.5 and available in an early research preview via waitlist. Opus 5.5 fast mode is open to everyone on the Claude Platform at $8/$40 with up to a 2.5x speedup. DeepSeek-V4.1-Flash-1017 shipped without a version-number bump and its benchmark chart lists Opus 5 and 5.6 Sol as comparisons; Mowshowitz's aside about how it came to exist points at distillation, and a commenter notes DeepSeek's new terse reasoning style resembles Astra's. Dean Ball is quoted estimating distillation explains roughly 30% of China's proximity to the U.S. on benchmarks by his prior, revised to about 50% after joining OpenAI. OpenAI, meanwhile, is expanding access to GPT-Rosalind for scientific research, bringing its models to Google Cloud, and offering GPT-Live-1.1 with new voices in multiple languages.
Muse Spark Open is now an open-weight model Meta calls the most capable U.S. open-weight model to date; it scores 39 on Artificial Analysis against an open-weight frontier of 46 held by MiMo-V2.6-Pro and Qwen-3.8, and MiMo is much cheaper to run. Nathan Lambert calls it his Super Bowl, cites the Interconnects Adoption Dashboard placing Spark at #8, and Mowshowitz credits the release with giving users an American open model for settings where Chinese weights are unwelcome, even run locally. He notes no major Chinese open-weight release landed between September 21 and October 16; Lambert cites compute restrictions and more internal lab-government safety discussion in China, while another theory holds that improved distillation defenses at Anthropic and OpenAI have the Chinese labs stuck.
Muse Wonder, Meta's flagship, gets a harsher verdict. Meta's marketing, including Alexandr Wang calling it the best model in the world for personal and professional agents, drew early praise from Eleanor Berger and a top ranking from Theo, but Kelsey Piper calls it narrower than current Claude and ChatGPT offerings and markedly worse at research and writing. Mowshowitz reports his reaction thread drew only two overnight responses and concludes Meta is making progress yet remains behind the frontier, with its reputation for overselling reinforced.
Benchmark notes: Epoch has reconstructed sped-up runs of SpireClimb, its Slay the Spire benchmark, where Ascension 0 win rates are near 100%; on Ascension 20, models beat the average player even among those who unlocked max difficulty. Benchmark creator Benjamin Ou suspects reinforcement learning on the game plus some generalization. HouseholdBench, which asks models to predict household spending, shows models cutting error by about 12% over an unchanged baseline, with open-weight models lagging.
Agent friction is a recurring theme. UPS, Delta, JPMorgan Chase, Charles Schwab and Amazon block AI agents, a set Mowshowitz expects to grow, making agents less useful. Over-refusals and permission prompts are the other barrier: he argues Claude auto mode asking for permission so often is actively less safe, because users stop reading the prompts, and his own rule against deleting anything outside source control pushed his Dot to extremes. Sharing a Muse conversation risks the link being indexed publicly. Patrick McKenzie's first multi-agent workflow left him arguing for an hour with a coordination process, and he came away convinced it is the future and that it will consume unlimited tokens.
On deepfakes, Mowshowitz argues they remain a limited practical problem despite political attention. New York City Council Speaker Julie Menin told a SpaceXAI representative at a hearing that Grok had produced a sexualized image of her and said she would follow up after unsatisfying testimony, also calling the representative's denial about a fired employee a lie. Stanford drew attention for race- and gender-swapping students in advertising imagery. Ethan Mollick frames the coming choice as adding frictions to services or removing them as agentic inquiries flood in, and founders report a surge of frivolous lawsuits, which Mowshowitz expects to keep rising.
Why it matters: The PHP disclosure is the item with immediate operational consequences, and teams running PHP anywhere in their stack should treat the disclosure-to-patch window as live exposure and inventory long-tail deployments that may never get patched. For product teams, the week's pattern is that agents are colliding with real-world gatekeepers: institutions are already blocking them, and permission-prompt design is a safety and usability decision, not a detail. On models, Meta's open-weight Spark is a credible U.S. option and Wonder looks oversold, so procurement decisions should rest on independent evaluations rather than launch claims, which is my inference from the reactions reported.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
It was the probably worst day in cybersecurity history.