cd /news/ai-safety/six-hours-and-a-stack-of-md-files · home › topics › ai-safety › article
[ARTICLE · art-170469] src=security.is.fucked.blog ↗ pub= topic=ai-safety verified=true sentiment=↓ negative

Six hours and a stack of .md files

Google Threat Intelligence Group reported observing a suspected financially motivated threat actor use an AI coding chatbot, a prompt and a set of agent instructions to plan, build and execute a mass credential harvesting campaign in less than six hours, stealing thousands of third-party credentials. GTIG said the attack ran from a victim's own cloud, and in another case a single leaked GitHub token let an attacker activate the victim's AI services and request more GPUs from the cloud provider; average darkweb marketplace prices for AI accounts including Claude, Gemini and Cursor are more than doubling in 2026. GTIG added it has not yet observed threat actors deploying fully autonomous pipelines against targets in the wild, describing instead a "gradual maturation of tradecraft and layering of AI capabilities.

by read5 min views1 publishedNov 16, 2026
Six hours and a stack of .md files
Image: Security (auto-discovered)

A few days ago, I posted:

Anyone can commit what used to be some of the most sophisticated and advanced cyber attacks ever done for less than it costs to buy a mcsmart meal at mcDonald’s

But after that, being the world’s biggest hater of misinfo, I went digging further to find out if my claim is true, and along the way, to find out how much... we.are.fucked(.blog)

Six hours #

Six hours is a lot of time or very little time depending on who you ask. When it’s a flight from one place to another that’s nothing. But when you are at war, that is insanely fast.

Google Threat Intelligence Group (GTIG), reports that they have been observing many cases of malicious use of AI models. And in one case, “a suspected financially motivated threat actor” used an “AI coding chatbot, a prompt, and a set of agent instructions to plan, build, and execute a mass credential harvesting campaign in less than six hours.” 1

So someone, maybe one guy, with a stack of .md files, in an afternoon, just... stole thousands of third-party credentials.

Let me say this again... a stack of .md files let the attacker operate at a scale and speed of large organizations with much much larger resources... 1 ... in just six hours.

Who paid for it #

Since the cost of premium model access and high-performance compute is one of the primary barriers for threat actors seeking to use AI for such missions. <sup>1</sup> What are they gonna do?

Steal the AI.

The six-hour attack was run from a victim’s own cloud. In another case, all it took was one leaked GitHub token, and then the attacker simply turned on the victim’s AI services and asked the cloud provider for more GPUs. We don’t actually know if the cloud provider scaled the demand and helped the attacker with their compromised account.

So it comes at no surprise to anyone that average darkweb marketplace prices for AI accounts (Claude, Gemini, Cursor) are more than doubling in 2026. 1

So currently the claim stands, as a McSmart meal costs more than what the attackers pay, which is close to nothing, because the victims get the bill.

What it can’t do yet #

GTIG say they have not yet observed threat actors deploying fully autonomous pipelines against targets in the wild, yet. 1

On the bright side it seems not every script-kitty can execute these attacks just yet. GTIG further claims that they observe a more “gradual maturation of tradecraft and layering of AI capabilities.” 1

This is not to say some amateur script-kitties aren’t trying to get in the game. With rapid prototyping for abusing existing bugs using fully AI-generated exploits.

In a recent-ish post by Anthropic they claim this is an open-weight problem <sup>2</sup> arguing that “attackers can’t readily access” the less-guarded versions of their models. Which are given only to vetted users. But GTIG’s own reporting shows adversaries are using BOTH commercial and open-weight models to turn disclosures into working n-day exploits. 1

But since the cheapest way to properly use these models at a proper scale needed for these attacks without getting caught is theft, another game needs to be played.

If it runs, it can be borked in to #

Prompt injection is not new, it feels like every other person on the internet has “Ignore all previous instructions” in their bio.

The real problem is that it is very easy to sneak something malicious to a personal agent of a dev, and usually agents are quite overscoped too with permissions to run wild. So a simple script in with a small instruction in .claude/, .cursor/, .vscode/, or whatever, is bound to do the trick eventually. 1

How do attackers hide malicious scripts? They place a block of text at the start of the malware asking for weapons instructions, purposefully, so that AI security scanners refuse to read the file and never look at the code underneath for the actual malware. Or at least this is the working theory by GTIG. 1

Meanwhile, on defense #

attacks got faster, a lot faster, but did defensive capabilities follow?

Somewhat. IBM asked 602 breached organisations, and there they found that more than 50% of them use AI agents to detect and contain, but only 18% use them to fix holes. 3

The difference matters because of who’s on the other side. That six-hour attack included an agent scanning for holes, non-stop, without anyone watching it. A specialist can fix a hole properly, but not in six hours, and not at 3am. Only 18% have put an agent on that job.

so the read on the situation is that defense... well it is falling behind.

We were never secure #

In September on Fedi I said, “Open source does not mean not obscure... people aren’t gonna spend their time combing through the code.” So many foss projects exist out there, with massive codebases. Of course things slip through.

The reality of why a lot of agents are making such big splashes is mostly from the scale at which they can process information. (1M context fits ~100,000 lines of code)

All the vulnerabilities were not secret, they were just not yet read by anyone.

This mostly hurts small indie devs, vibecoders or not, they are likely not committing equal resources to this as the attackers do.

And in the end, yes you can pull off a very sophisticated cyber attack with AI. For less than how much it costs to buy... anything at mcDonald’s, but are they “some of the most sophisticated and advanced cyber attacks ever done?” No, they are not, and it was an overstatement on my part.

stay safe,

micr0

Google Threat Intelligence Group, “GTIG AI Threat Tracker: From Prompting to Autonomy – The Evolution of Adversarial AI”, 8 September 2026. https://cloud.google.com/blog/topics/threat-intelligence/from-prompting-to-autonomy-the-evolution-of-adversarial-ai It covers Q2 2026, is limited to what Google and Mandiant can see, and gives no cost figures. 2. GLM-5.3 and the spread of advanced cyber capabilities https://www.anthropic.com/research/glm-5-3-and-the-spread-of-advanced-cyber-capabilities 3. IBM, “Cost of a Data Breach Report 2026”, press release of 29 July 2026 . Research by Ponemon Institute, sponsored and analysed by IBM, from 602 breached organisations between March 2025 and February 2026. IBM sells security products.

── more in #ai-safety 4 stories · sorted by recency
── more on @google threat intelligence group 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/six-hours-and-a-stac…] indexed:0 read:5min 2026-11-16 · —