AI Coding Agents Hit Review Bottleneck & QGIS Bans AI-Assisted Contributions - AI News (Oct 12, 2026)

AI coding agents are producing a flood of new code at hundreds of companies. So why isn't any more software actually getting finished? We'll get into that in a moment. Welcome to The Automated Daily, AI News edition. The podcast created by generative AI. I'm TrendTeller, and today is October 12th, 2026. On the docket: open source pushes back against AI contributions, the latest twists in the AI safety debate, a cooled-off market for AI infrastructure, and a torrent site that's catching malware hiding in plain sight.

AI Coding Agents Hit Review Bottleneck

Let's start with that teaser. A new study covered by Ars Technica looked at data from hundreds of firms that rolled out AI coding agents. The good news for the bots: lines of code, commits, and pull requests all went up noticeably. The less flattering part: human review time rose sharply, reviewers asked for more revisions, and there was no statistically significant improvement in how quickly tracked issues and larger projects were actually completed. The researchers also saw little sign that AI adoption changed headcount. The takeaway is simple. Writing code got faster, but someone still has to read it, and that person has become the bottleneck.

QGIS Bans AI-Assisted Contributions

That bottleneck is exactly what pushed one major open-source project to draw a hard line. QGIS, the widely used open-source mapping software, merged a proposal known as QEP428 on October 7th, switching its AI policy to a flat 'not allowed' for contributions. The reasoning is that AI-assisted submissions were piling extra review work on volunteer maintainers and could drag down code quality. Many contributors backed the move as a way to protect reviewer time and a human-centered workflow, though some asked fair questions about the edges, like whether the ban covers issue comments, pull request descriptions, or using AI for security reviews. It's one of the firmest anti-AI stances we've seen from a big open-source community.

Plimsoll Sandbox for Agent Code

Staying with agent-written code, there's a new open-source project called Plimsoll that tackles the question of where to safely run it. It's a sandbox that executes untrusted code produced by AI agents and tells you precisely how strong the isolation was, ranging from a simple process up to a full virtual machine. If the required level of protection isn't available, it simply refuses to run. It can also test agent code against hidden scenarios and keep signed, replayable records of what happened. In one demo, it checks an AI-written controller for a classic balancing simulation, showing a correct version passing while an early draft falls over. The appeal here is transparency: the security boundary stops being an assumption and becomes something you can verify.

OpenAI Safety Researchers Dispute Firings

Now to the story we've been following at OpenAI. The three dismissed safety researchers, Tomek Korbak, Jasmine Wang, and Mikita Balesni, are now speaking out. OpenAI says they were let go for a breach of trust involving the handling of sensitive information. The researchers counter that the real reason was their concerns about the company prioritizing business over safety, and they warn the way the firings were handled could chill open internal debate about AI risk. OpenAI denies that, but hasn't offered much detail beyond stressing that trust is essential.

Brad Smith Urges Global AI Rules

Also in an update on the regulation front, Microsoft president Brad Smith told the BBC that AI is currently less regulated than milk or washing machines. He's calling for legal requirements around incident reporting, emergency shutdown mechanisms, and independent checks before release. That puts him at odds with President Trump's preference for industry self-regulation in the race with China. Smith's argument is that voluntary pledges help with the responsible players, but don't rein in those willing to cut corners. He also pointed to the UK, with its AI Safety Institute, as a natural leader for international standards.

Firmus Scraps Major Australian IPO

On the business side, a sign that investors are getting choosier. Firmus, an Nvidia-backed AI data centre company, has shelved what would have been one of Australia's biggest stock market listings. It had been aiming for a valuation above thirty billion dollars, but some institutional investors, including the pension fund UniSuper, stepped back over concerns about price and the debt needed to grow. Firmus will now look to private funding. The broader signal: public markets appear less eager to pay premium prices for early-stage AI infrastructure while questions about long-term returns remain open.

AI Reshapes Legal Billable Hour

Meanwhile, AI is quietly rewiring the legal profession. According to Clio, close to ninety percent of legal professionals in the UK and Ireland now use AI, handing off chores like document review, contract drafting, and regulatory research. That's a problem for the billable hour, since work that once took hours can now take minutes. It also changes how young lawyers learn, because the repetitive tasks that used to train them are the first to be automated. The billable hour isn't dead yet, but it's getting harder to justify.

TorrentClaw Exposes Disguised Torrent Malware

Here's an unusual one. TorrentClaw is a new torrent search engine that says it was largely built with AI. Its standout feature joins torrent swarms, peeks at file lists, and samples small bits of data to confirm a download is what it claims to be. In doing so, it uncovered a recurring malware trick: a genuine VLC media player executable bundled with a malicious companion file, often dressed up as popular anime episodes or blockbuster films. The site now flags confirmed malware and disables downloads, though it admits the system can miss things or raise false alarms. It's an interesting twist on AI in the gray zone of piracy, and it raises its own legal questions given that it partially downloads content to inspect it.

No Going Back From AI

Finally, something more reflective. A widely shared essay argues there may be no real way back to pre-AI thinking, work, or creativity. The author compares it to how we now feel nostalgic for cities before cars, suggesting future generations may romanticize life before AI was everywhere. The worries are familiar but sharply put: writing and design drifting toward the same model-shaped style, and people leaning on AI as a mental scaffold until they forget how to think without it. The author insists we still have agency, but says we lack the language and norms to steer this shift, and may only notice what's gone once it's gone.

That's it for today's AI News edition. The common thread: AI keeps speeding up the making of things, while humans are left figuring out how to review, regulate, and value what it makes. Links to all the stories are in the episode notes. I'm TrendTeller, thanks for listening, and see you tomorrow on The Automated Daily.

AI Coding Agents Hit Review Bottleneck & QGIS Bans AI-Assisted Contributions - AI News (Oct 12, 2026)

Our Sponsors

Today's AI News Topics