AI Dev Brief #1: AI-found PHP bugs, stolen API keys, two new frontier models

Three things that mattered

1. AI found critical bugs in PHP. Patch now.

On 21 October PHP shipped emergency releases for two remote-code-execution flaws in the PHP core, CVE-2026-109618 and CVE-2026-109619, each rated 9.2 (critical). They affect every PHP installation that serves web requests, including PHP-FPM, Apache mod_php and FrankenPHP, and there’s no configuration workaround (Bitsight).

Branch	Update to
8.5	8.5.12
8.4	8.4.27
8.3	8.3.36
8.2	8.2.35
7.4, 8.0, 8.1	Out of support: source backports only; ask your OS vendor or host

CVSS v4.0 9.2 (critical) each. There is no configuration workaround; updating is the only full fix.

Source: Bitsight

The discovery story is the trend. OpenAI says the security firm Calif found one of the flaws using GPT-5.6 Sol, and that Wiz independently found and reported both; OpenAI coordinated the disclosure over the preceding two months (PPC Land). Exploit tooling was circulating by release day (Bitsight).

If you build software: run php -v everywhere PHP runs, including container base images, managed hosting and anything a vendor ships you. AI is shortening the time from bug to exploit on both sides, so how long a patch takes to reach production now matters more than it did.

2. An agent console leaked its users’ API keys

On 17 October OpenCode said attackers had taken provider API keys that 1.4% of OpenCode Console users had linked from Anthropic, OpenAI and Google. Some users reported unauthorised charges of up to $1,700, and some Google projects were suspended for hijacked-resource abuse (BleepingComputer).

If you build software: a key stored in a third-party tool is only as safe as that tool. Give each tool its own key with a spending cap, so one breach costs one key and a bounded bill. Rotate any key you’ve pasted somewhere you no longer use. Our essentials guide covers the rest.

3. Two new frontier models, at converging prices

Google made Gemini 4 Argon generally available on 13 October through the paid Gemini API, AI Studio and its Antigravity coding tool (The New Stack). The day before, Meta released Muse Wonder, its first frontier model, through the Meta Model API (TechCrunch).

Model	Input	Output	Notes
Gemini 4 Argon (until 30 Nov)	$2	$10	Introductory price; cached input 95% off
Gemini 4 Argon (from 1 Dec)	$4	$20	One 1M-token budget shared by input and output
Meta Muse Wonder	$4	$20	Meta Model API; Muse Wonder Full is not on the API

Sources: The New Stack, TechCrunch

If you build software: if you test Argon at the introductory price, budget for the doubling on 1 December. And note that Argon’s 1M-token limit is shared between input and output, so a long prompt leaves less room for the answer (The New Stack).

Releases and changes

• Open weights. Meta released Muse Spark Open on 15 October (The New Stack), and DeepSeek published V4.1 Flash 1017 under MIT (LM Studio). Read the licences before you build on either.
• Google Play, for one. Since 13 October, US Android users aged 18 and over can build an app in Google AI Studio and run it privately through Google Play. It isn’t listed publicly unless it passes Google’s review (TechCrunch).
• Supabase, 30 October. Existing Supabase projects stop automatically exposing new tables to the Data API. If an app or agent creates tables, it will need explicit grants (Supabase).

	Meta Muse Spark Open	DeepSeek V4.1 Flash 1017
Context	1M tokens	Up to 1M tokens
Licence	Muse Model License	MIT
Commercial use	Free in your own product; firms over $20M revenue need an agreement to offer it as a service; must show "Built with Muse Spark Open"	Permitted under MIT
Where	Hugging Face, the big three clouds, and hosted APIs incl. Groq, Fireworks, Together, OpenRouter	Hugging Face: deepseek-ai/DeepSeek-V4.1-Flash-1017

Sources: The New Stack, LM Studio model catalogue

By the numbers

Stack Overflow published its 2026 Developer Survey on 6 October, just before this period. The most common answer on trusting AI at work was “when I can easily verify it” (Stack Overflow).

1. When I can easily verify it48%
2. For many tasks, not important decisions16.3%
3. For low-risk tasks only16.3%
4. Not for most work tasks12.3%
5. Including important decisions6.6%

Stack Overflow 2026 Developer Survey, published 6 October. Respondents chose to take the survey, so treat it as a picture of engaged developers, not all of them.

Source: Stack Overflow

Our take (opinion)

The model news was about price and capability, but the two stories that cost people money were about keys and patches. As frontier models get cheaper per token, the costly mistakes move elsewhere: a leaked key with no spending cap, or a base image nobody rebuilt. For most teams, the most useful AI investment this fortnight wasn’t a new model. It was an hour spent capping keys and checking PHP versions.

Prediction check

None yet. The scorecard opens with our first quarterly outlook in January, and each prediction will have a date by which it can be checked.