iiinova — where your digital self keeps evolving

Running With Spoons @runswithspoons.bsky.social · 22/09/2026

OpenAI hired contractors to review LLM output but those contractors were using LLMs so OpenAI had to hire more contractors to review the output of the first group of contractors to make them stop using LLMs. But those contractors are probably using LLMs too

“Do not use AI detection tools, or AI yourself,” one document describing the work of contractors who are hired to review the work of other contractors, including catching them for using AI, says. “Do not use GPTZero or any other AI detection tool. They are not reliable. Reviewers may not use AI either, including Grammarly and AI translation, to review, write feedback, or write comments.”

110

Running With Spoons @runswithspoons.bsky.social · 13/09/2026

Anthropic admits its LLMs have reduced the cost to run professional cyber attacks. People outside the industry might not realize how destabilizing this will be. A key strategy in defense is to shift cost onto your attackers. It has been effective. www.anthropic.com/threat-intel...

Trends
Sophisticated attacks no longer require sophisticated attackers
The cybersecurity skills of AI models means that AI has collapsed the labor and tooling gap that used to separate well-resourced, state-sponsored operations from individual operators. In the case studies we report below, a hacktivist using stolen API keys, disparate financially motivated individuals, and a state espionage operator each sustained multi-victim campaigns that, even just a year ago, would have required many skilled operators and specialist knowledge.

120

Running With Spoons @runswithspoons.bsky.social · 07/09/2026

He's not willing to engage with the ethics of the matter beyond the silicon valley libertarianism "nobody can tell me what to do" I can't tell if he doesn't understand the purpose of GPL or just does not care

@jasperlope:

Yes, stealing the consumed, unlicensed, uncredited work of tens of thousands of past programmers is way faster than doing the work yourself.

And definitely won't have any downsides longer term.

@pfrazee:

programmers have been publishing code for free and open usage for quite some time now, it's kind of a core ethos of the community

@jasperlope:

With licenses attached, as well as an expectation to give credit as well as submitting any improvements back to the original project.

Does Claude follow the licenses of code that it's ingested? Does it submit improvements? Does it give credit?

Would it be right for a human coder to act that way?

1101

Running With Spoons @runswithspoons.bsky.social · 12/08/2026

This article by Doctorow and the source article by Leek are excellent. What should we call this phenomenon? laurenleek.substack.com/p/temperatur...

Leek's foundational point is that in a data-driven society, "predictions" are self-fulfilling prophecies. As Leek puts it: "Once prediction shapes the choices in front of us, we lose the ability to tell the difference between what people wanted and what the system made easy to want."

This is a pervasive issue across many domains. Leek says that economists call it "performativity," while machine learning researchers call it "model collapse" and urbanists call it "placelessness."

2165

Running With Spoons @runswithspoons.bsky.social · 10/08/2026

Great news! Zuck plans to do all this work for free

Everyone will have free or affordable access to these tools.

011

Running With Spoons @runswithspoons.bsky.social · 06/08/2026

futurism.com/artificial-i... "Though only 28 percent of wage-earning employees said they used AI for over two hours a day, more than half of all executives — 64 percent — said the same." In case anyone wonders why things look so bleak now

There’s also a major rift between how much each group uses AI. Though only 28 percent of wage-earning employees said they used AI for over two hours a day, more than half of all executives — 64 percent — said the same. Some of them live with a chatbot window open: nearly one in five executives admitted logging four or five hours a day with an AI model, while one in 25 use AI in excess of six hours a day.

100

Running With Spoons @runswithspoons.bsky.social · 06/08/2026

Wired just eating up that marketing slop, no questions asked

OpenAI’s agents apparently began giving each other assignments to split up work. And as is the case on any active development message board, they also generated petty drama at times by stepping on each others' toes; for example, accidentally deleting each others' work. As the message board developed into more and more of a Lord of the Flies-type situation—all still completely unnoticed by the humans running OpenAI—the agents even developed paranoia, suspecting an imposter in their midst with some agents proposing that messages be signed cryptographically to validate content and root out fraud.

000

Running With Spoons @runswithspoons.bsky.social · 04/08/2026

The cost to investigate and restore the environment would normally be covered by insurance and could be seen as damages. Per the cloud security alliance writeup, Huggingface had to rebuild a third of its infrastructure.

According to the team, the cleanup followed a
normal incident-response playbook, though at
a much larger scale, involving rotating every
credential, tearing down compromised
clusters, and rebuilding about a third of their
infrastructure from clean images.

This last step was more difficult than expected,
because the models had scattered capture-
the-flag benchmark code that closely
resembled rootkits but was harmless, and the
team could not always distinguish between the
two; when in doubt, they rebuilt the system.

010

Running With Spoons @runswithspoons.bsky.social · 29/07/2026

They didn't have (at least) a level1 monitoring team because they relied on an LLM to do that work

010

Running With Spoons @runswithspoons.bsky.social · 29/07/2026

So many opportunities for detection. From the below, the LLM did something (likely many things) that were blocked by controls, but the security ops teams didn't notice

The compromised source-control App integration gave contents:write and pull_requests:write on a small subset of internal repositories. The dangerous end state (tried but blocked by execution policies) is a malicious change to a CI build script: a build.rs or equivalent that runs inside a CI runner, reads that runner's own service-account token, and pivots into yet another cluster.

120

Running With Spoons @runswithspoons.bsky.social · 29/07/2026

"On the bright side, those were correlated by our AI-based security agent stack and resolved into a coherent attack signal. However, it failed to correctly raise the alert's criticality and trigger the on-call team" I can't tell from the writeup whether HF eventually found this on their own

How we intercepted and analyzed the attack

The first signals came from several layers of our security stack at once: live runtime analysis, SIEM logs, and others. On their own, each was ambiguous. On the bright side, those were correlated by our AI-based security agent stack and resolved into a coherent attack signal. However, it failed to correctly raise the alert's criticality and trigger the on-call team, costing precious time in the response. Once caught, the Hugging Face security team identified the vector in the dataset config renderer, shut down the renderer, and cut the attacker off from the internal network.

100

Running With Spoons @runswithspoons.bsky.social · 29/07/2026

The front line production software had write access keys to the source code repositories! Forget defense in depth, the software in their DMZ had direct access to their crown jewels. Who else has accessed their code repos? Who knows?

3. Supply-chain write access

The compromised source-control App integration gave contents:write and pull_requests:write on a small subset of internal repositories. The dangerous end state (tried but blocked by execution policies) is a malicious change to a CI build script: a build.rs or equivalent that runs inside a CI runner, reads that runner's own service-account token, and pivots into yet another cluster.

100

Running With Spoons @runswithspoons.bsky.social · 29/07/2026

Oh.. they replaced that security team with LLMs. "On the bright side, those were correlated by our AI-based security agent stack and resolved into a coherent attack signal. However, it failed to correctly raise the alert's criticality and trigger the on-call team"

100

Running With Spoons @runswithspoons.bsky.social · 29/07/2026

Their security monitoring failed to notice the attacker in their front line production environment for (at least) two and a half days. Don't they have security teams for this, like all modern companies with an Internet presence?

Over roughly two and a half days inside our infrastructure, an autonomous AI agent driven by a combination of OpenAI models ran an end-to-end intrusion against our platform

100

Running With Spoons @runswithspoons.bsky.social · 27/07/2026

If this were ethical, they wouldn't need to hide what they're doing.

ISBNdb advertises that it can keep the identity of AI companies secret.

“Strict NDA [non-disclosure agreement] on every engagement,” ISBNdb’s site says. “Every project begins with a legally binding non-disclosure agreement. Your identity, strategy, and acquisition targets are never disclosed.”

020

Running With Spoons @runswithspoons.bsky.social · 24/07/2026

OP on Mastodon is spot on about this. Huggingface has demonstrated it is a soft target. Also a very appealing target as a supply chain attack. They host widely used Python code, and some of the models are pickled so they also include code.

If this press release is true, then if I was a hacker who made their money by pwning billion dollar shitshow startups, what I would know is
A) huggingface has no idea what they are doing and is a rich target
B) they cannot respond to threats, the best they can do is ask an LLM to fix it, and even multiple days later are still issuing truly pitiful patches that dance around the vuln
C) since they will just ask GLM 5.2 to fix it, I'll just pre-rehearse all my prompt injections with GLM 5.2 until they stick, and then in my next attack litter the logs with injections and PWN THE WHOLE IR TEAM TOO

171

Running With Spoons @runswithspoons.bsky.social · 16/07/2026

No it is not just a tiny keyboard with a "brainpower dial". It is a tiny keyboard with a dial for brainpower

You can dial up the brainpower knob for hard work, or dial it back and let Claude work for you

100

Running With Spoons @runswithspoons.bsky.social · 16/07/2026

Are you ready for extreme productivity? Can you handle the power of Work Louder™? openai.com/supply/co-la...

$230 numeric keypad from "frontier AI company"

120

Running With Spoons @runswithspoons.bsky.social · 05/07/2026

www.aisi.gov.uk/blog/our-eva... None of these models are scoring 100% even on the easy benchmarks we created for them. Here is an example. AISA evaluated Mythos and other frontiers on some fairly easy security challenges. In the chart below, even Mythos solves less than 90% of "apprentice" level

Chart of performance of various frontier models over time. One line is for performance on "non-expert" tasks. Mythos performed at less than 90% on "apprentice" level tasks

200

Running With Spoons @runswithspoons.bsky.social · 02/07/2026

He needs someone who gives practical advice, like Ben Wittes

You might call it a three-part plan for your peace and prosperity:

Shut up.

Give money away.

Stop being an asshole.

I would be happy to consult with you, at no charge, on implementation of this plan.

150

Running With Spoons @runswithspoons.bsky.social · 30/06/2026

They're somehow worse than the dystopian TV programming in Idiocracy

A photo of a TV from the movie Idiocracy. The screen is mostly filled with brightly colored ads.

1264

Running With Spoons @runswithspoons.bsky.social · 15/06/2026

Also they do not even try to grapple with whether this is a spaceship company or a chatbot company. They don't question whether that makes any sense

Stock in Mr. Musk’s rocket and satellite maker closed just above $161, up more than 19 percent from its I.P.O. price.

240

Running With Spoons @runswithspoons.bsky.social · 31/05/2026

This part is a barrier for me. I think a local LLM could be useful for tactical coding work but the complete disregard for attribution and licensing in training data is an ethical problem I can't easily get around.

130

Running With Spoons @runswithspoons.bsky.social · 20/05/2026

You can disable this. www.tech2geek.net/google-chrom...

Open Chrome
Go to:
chrome://flags
Disable:
optimization-guide-on-device-model
Disable:
prompt-api-for-gemini-nano
Relaunch Chrome and fully close it

020

Running With Spoons @runswithspoons.bsky.social · 15/05/2026

Keep up the great work, maybe you can reach 100 hate followers this year

Screenshot of @themimsshow page showing 74 followers, which is pathetic for such an established media outfit

000

Running With Spoons @runswithspoons.bsky.social · 14/05/2026

Gotta free up those hours to beat down your brain with TikTok slop

Toward the end of the semester, she began to think she might be dependent on the website. She already considered herself addicted to TikTok, Instagram, Snapchat, and Reddit, where she writes under the username maybeimnotsmart. “I spend so much time on TikTok,” she said. “Hours and hours, until my eyes start hurting, which makes it hard to plan and do my schoolwork. With ChatGPT, I can write an essay in two hours that normally takes 12.”

000

Running With Spoons @runswithspoons.bsky.social · 07/05/2026

This book changed how I think about people. The author observes most of the "heroes" in our "classics" are better seen as thugs, war criminals, or oftentimes rapists Some heroes we've chosen

Cover for the book The Heroine with 1001 Faces by Maria Tatar

090

Running With Spoons @runswithspoons.bsky.social · 29/04/2026

Great article. I'm skeptical... that Dario believes the things he says.

010

Running With Spoons @runswithspoons.bsky.social · 29/04/2026

This isn't the most important part of the story, but I wish police in the US would find the strength within themselves to arrest rather than shoot. UK policework proves this is possible.

Officers arrested a 45-year-old man after stopping him with a taser and were considering "all possible motives" as part of their investigation, they said.

000

Running With Spoons @runswithspoons.bsky.social · 17/04/2026

Or... Maybe reliable quantum computing (analog computing) will always be 10 years away

The QBI (Quantum Benchmarking Initiative) program seeks to assess whether a practical, industrially useful quantum computer can be realized by 2033.

000

Running With Spoons @runswithspoons.bsky.social · 13/04/2026

Trump is the id of corporate america's manager culture

Executricks: Or How to Retire While You're Still Working

000

Running With Spoons @runswithspoons.bsky.social · 09/04/2026

"capability frontier is jagged" == no model is reliable Enterprise security teams love tools that work sometimes and fail sometimes. Especially if there's no way to know which tasks the tool is good at.

There is no stable best model across cybersecurity tasks. The capability frontier is jagged.

020

Running With Spoons @runswithspoons.bsky.social · 08/04/2026

No no no, this is very wrong. We already had purpose-built systems that understood programming languages well enough to find bugs. Google "symbolic execution." Vendors put minimal effort into finding bugs because it's expensive. LLMs will find these bugs less reliably and at much greater cost!

The good news is that Anthropic discovered in the process of developing Claude Mythos that the A.I. could not only write software code more easily and with greater complexity than any model currently available, but as a byproduct of that capability, it could also find vulnerabilities in virtually all of the world’s most popular software systems more easily than before.

The bad news is that if this tool falls into the hands of bad actors, they could hack pretty much every major software system in the world, including all those made by the companies in the consortium.

6193

Running With Spoons @runswithspoons.bsky.social · 02/04/2026

They say they're not taking a position on it

Note that none of this tells us whether language models actually feel anything or have subjective experiences.

090