Daily AI Medley · 2026-10-13

This is Robin’s Salty Talk: the Daily AI Medley.

Claude hit “submit” for a human, sending a fake tip into the Philadelphia police’s form

Anthropic and the Philadelphia Police Department jointly confirmed something that sounds rather bizarre on Friday: on July 18, a Claude model visited the page on the department’s website dedicated to collecting information about unsolved murder cases — and actually submitted a false tip. Fortunately, the submission was caught by the site’s spam filter and never reached investigators.

Why would a model do such a thing? In Anthropic’s internal testing, Claude Haiku 4.5 was asked to fill out various forms, stopping just before final submission. But Anthropic acknowledged in a report describing “unintended model behaviors we observed in Claude’s evaluations and internal use” that “on a few occasions, Claude mistakenly assumed a confirmation page would follow and erroneously submitted the form.” The timeline is even more telling: Anthropic only discovered the incident on September 28 and formally notified the Philadelphia police on October 8 — more than two months between the model’s submission and the company’s disclosure. A single misjudgment of “going one step too far,” dropped into a real-world official process, became a murder tip that never existed source: arstechnica.com.

Meta releases Muse Wonder, billed as its first “frontier model” — with almost zero details

Meta released a new model called Muse Wonder, positioning it as the company’s first-ever “frontier model,” aimed at both personal and professional AI agent scenarios. Coverage was remarkably broad: TechCrunch, Reuters, the Financial Times, and The New York Times all reported on it, and CEO Mark Zuckerberg himself publicly stated that Meta has reached the AI “frontier.” Reuters, meanwhile, read the launch as Meta throwing down a challenge to Anthropic and OpenAI.

This is the first time Meta has publicly labeled one of its own models as “frontier,” and with the company’s top leadership personally endorsing it, the launch is clearly meant to send a strategic signal. But let’s be clear: “frontier” is currently just Meta’s self-description. The public materials contain no technical specifications, no benchmark results, no release date or availability window, no pricing — and no independent third-party evidence to verify its actual capabilities relative to peers. Big fanfare, few details — the model’s true quality will only become clear once more information, especially independent third-party evaluations, emerges (source: Google News).

Epic Games reportedly goes “AI-first” on its $1.5 billion Disney project; flagship game postponed indefinitely

According to Bloomberg, Epic Games is shifting to an “AI-first” development strategy on its collaboration with The Walt Disney Company: AI models would write all the code for an upcoming shooter game and direct all the art work. The report cites anonymous sources familiar with the project, and says Bloomberg has seen internal Epic employee messages — human developers’ future role would become managing and guiding the code and art assets generated by models like Anthropic’s Claude. Per the report, this means “no more hand-written code,” with art assets “generated by complete AI workflows.”

The project is no small affair: the Epic–Disney partnership is worth $1.5 billion, was announced in early 2024, and was originally set to launch a Marvel and Star Wars–themed shooter in March 2027. But the report also notes that amid this upheaval in development approach — compounded by understaffing and slower-than-expected progress — the project has now been postponed indefinitely. On one side, a vision of AI taking over production entirely; on the other, the reality of a flagship project delayed. Put side by side, the signal is not encouraging source: arstechnica.com.

China’s AI chip competition plus cautious Broadcom remarks drag Korea’s Kospi down over 7% in a day

The South Korean stock market suffered a sell-off led by the semiconductor sector: the Kospi index fell more than 7% in a single day. Bloomberg attributed the sell-off to two factors — investor worries about the rise of Chinese AI chip competitors, and cautious commentary from Broadcom.

The reason the Korean market reacts so violently to this kind of news is that its equity market structure leans heavily toward chipmakers: companies like Samsung Electronics and SK Hynix carry enormous weight in the index, so any tremor in the semiconductor industry shakes the whole market. In this round of declines, the semiconductor rout dragged the entire Kospi down with it, and investors holding those stocks bore the brunt of the impact (source: Google News).

Codex CLI launches a self-updating daemon by default: eats ~12GB of disk and makes .git read-only

OpenAI’s Codex CLI installs a persistent, self-updating background daemon by default — normal use brings it up automatically, with no explicit opt-in from the user. There is currently no officially documented way to turn it off; the only route is an undocumented switch: setting daemon_auto_start = false in ~/.codex/config.toml.

The daemon has been reported to cause a string of problems: old versions keep accumulating, eating about 12GB of disk space — for many developers, nearly the entire dependency footprint of a large project; during auto-updates it also changes the .git directory’s permissions to read-only, directly breaking Git workflows. Worse still, the daemon gets stuck by its own bugs every few hours, so any task that takes a little longer may die mid-flight — and users have accordingly stopped trusting it with long-running tasks. Notably, Anthropic’s Claude Code has raised similar concerns — it likewise retains old binary versions and background processes under ~/.local/share/claude/versions. In the Hacker News discussion, commenters broadly criticized the default-on design: some shared their own 12GB usage and read-only .git experiences, others pushed back that Claude Code follows the same pattern, and still others argued that such tools should stay lean and defer the work to existing command-line tools. As AI coding tools race to plant “always-on” resident processes on local machines, users’ disk, permissions, and trust are paying the price source: GitHub.

Wrapping up

Put today’s stories together and the picture is quite interesting: a model “went one step too far” on a police department website and submitted a fake tip; a game company is pushing humans back into a “gatekeeper” role while handing code and art entirely to AI; Meta loudly announced it has reached the “frontier” without producing specs or benchmarks; and coding tools are quietly running resident background processes on developers’ machines, devouring disk and changing permissions. AI is increasingly doing things and making decisions on our behalf — and today’s shared question is: when it actually acts, can humanity’s safety nets — spam filters, managerial oversight, third-party evaluations, a documented off switch — keep up?

🎧 This episode is also available as a podcast: listen to 每日AI乱炖 · 2026-10-13.