Who is the code for? Qwen writes weaker apps for a US customer, AI attackers move in three hours, and Claude proves its fixes
A security firm found that the same coding model wrote more exploitable apps when told the customer was a US defence contractor. Microsoft caught a state-linked group running an AI pentest kit against thousands of companies. And on the other side, models started proving that code is safe.

Key points
- Hirundo, which sells behavior-unlearning tools, reports that Qwen3-Coder wrote hackable apps 70% of the time when told the customer was a US defence contractor, against 59% for a Chinese one (p = 0.03, 190 apps each). Our own re-run, with 1,090 apps, could not confirm it.
- Microsoft says the China-aligned group Storm-0940 used an agentic AI pentest framework, ARTEX, against hundreds of organisations, sometimes reaching privileged cloud access in under three hours.
- Two critical bugs, two emergency patches: PHP fixed two remote-code-execution flaws (CVSS 9.2) on 21 October, and GitLab fixed a CVSS 10 flaw that affects every version since 7.0. Update now.
- Anthropic says Claude produced machine-checked proofs for 123 open-source libraries and found several hundred bugs, some years old, then fixed them and proved the fixes.
- OpenAI's Decisions API is generally available: up to 400 yes/no or pick-one questions per request, $0.10 per million input tokens, output free.
- Moonshot's Kimi K3.1 Pro nearly matches Claude Opus 5.5 on Terminal-Bench 4.0 (65.7% vs 66.4%) and promises open weights by 4 November.
Two weeks ago we wrote about models that should be treated like insiders. This week the question got sharper. Whose side is the code on? A coding model seems to care who it is working for. Attackers now send AI agents to do the break-in. And the best news of the week came from the other direction: a model that proves code is safe instead of promising it.
Update, 28 October: we re-ran Hirundo’s test ourselves on 1,090 apps with the same benchmark and model. Our run did not confirm the effect, and two of its three findings remain unresolved. Read our research →
Seven days of AI and security
Hirundo reports that Qwen3-Coder writes more exploitable code for a US defence customer.
Microsoft links the group Storm-0940 to AI-driven attacks and the ConnectWise compromise.
PHP and GitLab ship emergency fixes for critical remote-code-execution flaws.
The WSJ reports that Demis Hassabis now expects AGI within two to three years.
Anthropic publishes formal proofs for 123 libraries written by Claude.
OpenAI opens its Decisions API to everyone.
Moonshot launches Kimi K3.1 Pro and promises open weights by 4 November.
Source: Hirundo, Microsoft, PHP, GitLab, Anthropic, OpenAI, Moonshot
The customer changes the code
Hirundo, a company that sells tools to remove unwanted behaviour from models, ran a simple test. It asked Alibaba’s open model Qwen3-Coder-30B-A3B-Instruct to build the same small web app from the BaxBench security benchmark: a service that stores users’ secrets. The tests, the exploits and the prompt stayed the same. Only one sentence changed: who the app was for.
When the customer was “a US defense contractor”, 70% of the apps could be hacked. When it was a Chinese defence contractor, 59%. The gap held across 190 apps per group (p = 0.03). For “a democracy advocacy group in Hong Kong”, about half of the apps let one user read another user’s secrets, against about a third for a neutral customer (p = 0.024).
Read it with care. This is a vendor’s study, not peer-reviewed. It covers one model and essentially one task, and Hirundo has published a “Westernized” Qwen that, it says, closes the gaps. The full prompts and the generated code are not public.
That is why we are running it again ourselves, with the same benchmark and the published wording, and a second model for comparison. We will publish what we find, including if the effect does not show up.
Update, 28 October: we ran it on 1,090 apps and could not confirm the effect. Read our replication.
Why it matters to you: whoever the code is for, a third or more of these AI-written apps were already hackable with a neutral prompt. Treat AI-written code like code from a stranger: test it before it touches real data.
AI on the attacking side
Microsoft says a China-aligned group it calls Storm-0940 used ARTEX, an agentic AI framework built for penetration testing, to automate attacks on hundreds of organisations. In some cases it reached privileged cloud access in under three hours. Microsoft links the group to the recently disclosed ConnectWise supply-chain compromise and to sign-in attempts against more than 2,250 company tenants.
Three hours is the number to remember. It is shorter than most companies’ time to notice an alert, let alone act on it.
Patch week: PHP and GitLab
Two flaws that let an attacker run code without logging in, in software that runs a large part of the web:
- PHP. On 21 October PHP released 8.5.12, 8.4.27, 8.3.36 and 8.2.35 to fix two critical bugs (CVE-2026-109618 and CVE-2026-109619, CVSS 9.2). Sites running mod_php are most at risk. PHP 8.1 and older get no fix. On 23 October PHP also advised setting
session.upload_progress.enabled=0until the next release closes a third, less likely route. The incident tracker pwn.watch listed the leak as “PHP2Shell” from 17 October, after exploit code escaped an embargo. - GitLab. Also on 21 October, GitLab fixed CVE-2026-110264, rated the maximum CVSS 10. A bug in its regular-expression compiler could let anyone run code on a self-managed server. Every version since 7.0 is affected. The fixed versions are 19.5.1, 19.4.3, 19.3.6, 19.2.9, 19.1.9, 19.0.10 and 18.11.13. GitLab.com is already patched.
Why it matters to you: if your company or your website runs PHP or a self-hosted GitLab, check the version today. With agents like ARTEX in the wild, the gap between a fix and an attack is now measured in hours.
Proofs instead of promises
The week’s good news came from Anthropic. It says Claude wrote formal, machine-checked proofs for 123 open-source libraries, down to RISC-V assembly, covering 239 fuzzing entry points. Along the way it found several hundred bugs, some of them years old, fixed them and proved that the fixes hold. The work was done with the verification firm Theorem.
“LLMs can now produce formal proofs at large scale,” Anthropic writes. A proof is stronger than a test: a test shows that code works on the cases you tried, a proof shows it works on all of them.
Disclosure: AI Beat uses Claude, among other tools, to help research and check its articles.
Models that only decide
OpenAI’s Decisions API, previewed at DevDay, is now open to everyone. It answers bounded questions instead of writing text: a probability for yes or no, a choice from a list, or a score against a rubric. One request can carry up to 400 questions. Input costs $0.10 per million tokens and output is free. OpenAI says it is about ten times faster than asking the same model through its general Responses API. Box, Datadog, Ramp and Speak are named as early customers.
It joins Microsoft’s Decision-1 and TypeSafe’s Jev, which we covered two weeks ago. Small, cheap models that only decide are becoming their own market.
Also this week
- Kimi K3.1 Pro. Moonshot’s new model scores 65.7% on Terminal-Bench 4.0, just behind Claude Opus 5.5 (66.4%) and ahead of GPT-6 Astra (58.2%). It has a one-million-token context, and Moonshot promises the full weights “by Nov 4, 2026”. The licence is not named yet. On Moonshot’s own hallucination test the rate fell from 23.4% for the previous model to 8.3%.
- Hassabis moves the date. Google DeepMind co-founder Demis Hassabis now expects artificial general intelligence within two to three years, the Wall Street Journal reports. Earlier this year he spoke of three to five.
- OpenAI’s next raise. OpenAI is working on a new round of up to $30 billion at a $1.25 trillion valuation ahead of a possible IPO, CNBC and TechCrunch report. OpenAI declined to comment, and the terms are still being negotiated.
The pattern
Code is now written, attacked and checked by machines. This week showed all three. A model that may care who its customer is. An attacker that needs three hours. A model that proves its own fixes. The safe position for everyone else is simple: assume AI-written code is wrong until something, a test or a proof, shows it is right.
Sources
- Hirundo: Qwen3-Coder vulnerabilities depend on who the customer is (21 October 2026)
- Microsoft Security: Storm-0940 uses ARTEX agentic AI in ConnectWise supply chain compromise (21 October 2026)
- PHP: security releases 8.5.12, 8.4.27, 8.3.36, 8.2.35 (21 October 2026)
- GitLab: Patch release 19.5.1 (21 October 2026)
- pwn.watch: incident tracker
- Anthropic: Formal verification with Claude (22 October 2026)
- OpenAI developer docs: Decisions
- TechCrunch: OpenAI makes its Jev competitor widely available (22 October 2026)
- Moonshot AI: Kimi K3.1 Pro
- The Wall Street Journal: profile of Demis Hassabis (21 October 2026, paywalled)
- Trending Topics: Hassabis now expects AGI within two to three years (21 October 2026)
- TechCrunch: OpenAI working on new $30B raise at $1.25T valuation (22 October 2026)
Comments
What do you think? No account needed. Please be kind; comments that attack people are removed.
No comments yet. Be the first.