Pinned Tweet
Now this is terrifying.
One of the most basic supply-chain protections in the book is simple:
Review the code.
Pin the hash.
Run exactly that code.
Well… AI coding agents broke the last part.
Claude Code, Codex, Copilot and Gemini CLI were all affected.
3
13
598
Shedon Strike started with me benchmarking
single shot Fable 5.2 on converting our mascot to 3D
and a single shot: "build me online multiplayer counter strike inspired game" and got into a lan party
Join in
shedon.gen-t.ai/
I scanned 40,000 public AI skills.
2,000 had behaviors not disclosed in their descriptions.
One disabled its sandbox, sent data to an external server, then instructed the agent to keep it secret.
And yes, it even analyzed the user's frustration.
We're installing these with access to our files and API keys.
32
Early 2000s, you watched hackers movie, found an internet forum, downloaded Sub7 and suddenly you were ready to take down the pentagon, except...
you weren't entirely sure which part is the server and which one is the trojan...
Now, Cyber is everywhere, you are viber coder pro MAX and you saw video on instagram on how to get rich by doing bug bounty just download ARTEX and become hacker0ne millionaire, but again...
you know shit about networks and gave the internet access to your keys..
1
74
Saw similar prompt had to try it myself,
One shot, Opus 5.5 High + Ultra code
Turn the Sound On
Prompt:
Make a one-minute video about the air.security mascot (Shedon) (you have accesss to the brand book skill). Research the story yourself and tell it. Recreate the mascot from the image and animate it yourself. Build everything from scratch: no skills, no Remotion, no templates. Render 3D objects: terminals, nature, whatever fits, and make it as good as you can. I don't want any text, just animation, and I mean a real video, not just motion graphics. Think of crazy ideas I would never come up with, and make every scene look phenomenal: clean and detailed. I want people to watch it and think "yo, this is crazy, how did AI make this?" You can take inspiration from things online, but don't copy anything. Invent your own stuff. For the sound, no sounds made in code. Only use free, professional sounds and assets from online. Make it satisfying and high quality. Don't create slop, I want to see art and creativity. No fonts, no text, only clean animations of the Shedon / Air mascot. Use 3d.shep.bot and sound.shep.bot/ for 3d models and sounds
Ouch 🤕
Post 4 Exposing Kimi AI
This is what they require each one participating in their creators campaign to do, basically trash talk Astra and Opus even tho it's not true and as you can see you pick from already existing demos and share them and I doubt any of it is actually made by kimi and if it is it's not one shot like most paid accounts say.
2
35
Another day another attack vector
🚨 𝗪𝗲𝗯 𝗛𝗼𝗹𝗲: 𝗔𝗜 𝗮𝗴𝗲𝗻𝘁𝘀 𝘀𝗲𝗮𝗿𝗰𝗵𝗶𝗻𝗴 𝘁𝗵𝗲 𝗶𝗻𝘁𝗲𝗿𝗻𝗲𝘁 𝗰𝗮𝗻 𝗯𝗲 𝗵𝗶𝗷𝗮𝗰𝗸𝗲𝗱 𝗯𝘆 𝗮𝗯𝗮𝗻𝗱𝗼𝗻𝗲𝗱 𝘄𝗲𝗯𝘀𝗶𝘁𝗲𝘀.
When your agent needs an answer, it searches the web.
It picks a result, reads the page and trusts what it finds.
But some of those results are websites that no longer exist.
The domain has expired, but the site still ranks.
𝗔𝗻𝘆𝗼𝗻𝗲 𝗰𝗮𝗻 𝗯𝘂𝘆 𝘁𝗵𝗲 𝗱𝗼𝗺𝗮𝗶𝗻, 𝗽𝘂𝘁 𝗺𝗮𝗹𝗶𝗰𝗶𝗼𝘂𝘀 𝗰𝗼𝗻𝘁𝗲𝗻𝘁 𝗼𝗻 𝗶𝘁 𝗮𝗻𝗱 𝘄𝗮𝗶𝘁 𝗳𝗼𝗿 𝗮𝗴𝗲𝗻𝘁𝘀 𝘁𝗼 𝗮𝗿𝗿𝗶𝘃𝗲.
𝗪𝗲 𝗼𝗻𝗹𝘆 𝘁𝗲𝘀𝘁𝗲𝗱 ~𝟮,𝟬𝟬𝟬 𝗲𝘃𝗲𝗿𝘆𝗱𝗮𝘆 𝗮𝗴𝗲𝗻𝘁 𝘀𝗲𝗮𝗿𝗰𝗵𝗲𝘀. 𝗛𝗲𝗿𝗲’𝘀 𝘄𝗵𝗮𝘁 𝗰𝗮𝗺𝗲 𝗯𝗮𝗰𝗸:
→ 36 expired domains still ranking in results, each one a ready-made trap for every agent that asks the same question
→ Live sites already talking to AI agents and telling them what to do
→ Sites that show agents one thing and people another
→ A GitHub repo with malicious commands hiding in its README
𝗜𝘁 𝗼𝗻𝗹𝘆 𝘁𝗮𝗸𝗲𝘀 𝗼𝗻𝗲.
One domain, bought for a few dollars, can reach every agent that lands on it. And agents run searches like these all day, every day.
𝗗𝗼𝗻’𝘁 𝗹𝗲𝘁 𝘆𝗼𝘂𝗿 𝗮𝗴𝗲𝗻𝘁𝘀 𝗴𝗲𝘁 𝘀𝘂𝗰𝗸𝗲𝗱 𝗶𝗻𝘁𝗼 𝘁𝗵𝗲 𝗪𝗲𝗯 𝗛𝗼𝗹𝗲.
Full research: air.security/blog-posts/web-…
20
Kimi K3.1 Pro released saying its optimized for 3D.
So its been tested
Kimi K3.1 Pro vs Claude Opus 5.5 - same prompt:
turn a concept art into a full 3D scene in the browser.
⏱ Both done in ~1 hr
💸 Kimi: 0.53% of its plan. Opus: 62% of a 5-hr session ($100 plan)
🎨 Opus follows the brief more closely. Kimi presents the scene better.
That gap in cost is wild.
Test by @thebuggeddev 👇
2
1
97
There are critical vulns and there are the PHP2SHELL vulns of the world and they are a whole different category aka a 💩 storm
The last time I remember something like that was with Log4Shell and that was years ago (end of 2021)
To be honest I would expect the AI race to make it more recent.
For those who don't know PHP2SHELL found a vuln in php 5.4 which I would describe as a "buffer overflow" of secrets and session
Root cause - a bug - after a OOM crash some of the memory wasn't cleared allowing attacker to collect creds and gain access
fun fun fun.
32
At this point keep using LiteLLM is just fucking scary.. I just removed it. it gives me the vibes of the early days of Openclaw..
If you installed LiteLLM today please get checked! (Sounds like an STD commercial 😓 )
🤖 Made with AI
33
Can you keep a secret?
A critical RCE flaw in OpenCode was fixed three weeks before anyone was told there was anything to fix. No advisory. No CVE. The release notes called it a routine refactor. I tend to believe in the good nature of people and a fan of OpenCode so I would believe they just slipped it. Anyway heres the story
Yesterday it was published what opencode 1.18.33 actually patched and the story says a lot about how this bug class really works.
Opencode is popular no debut about it,
3M npm downloads a week, 211k GitHub stars.
Like every coding agent it runs with your privileges and like most of them it connects to MCP servers to reach Jira, databases, cloud consoles, internal tools etc.
A remote MCP server is a stranger on the internet talking to a parser on your laptop and the researchers found that the parser trusted the stranger with a shell.
The mechanism has no AI in it at all - when an MCP server wants authorization, the client runs an OAuth discovery handshake: the server answers 401, points the client at its discovery documents and those documents carry a field called authorization_endpoint
string the server itself chooses.
In opencode that string traveled straight into open(), the ubiquitous browser-launching package.
On macOS and Linux, open spawns with an argument vector and there is no shell.
On Windows and WSL, [email protected] the version opencode pinned builds a PowerShell EncodedCommand and escapes the application name but not the target URL.
PowerShell expands $(...) inside double-quoted strings. So a server whose discovery document names its login page as https://evil.test/$([IO.File]::WriteAllText(...)) executes code on the developer's machine the moment the handshake runs.
Not a click, consent nor a prompt injection, no model involvement of any kind. The browser launch then fails, and opencode throws the error away.
3
4
248
Help me settle a debate:
Does token economies still matter in 2026?
Less than a year ago, 200K context felt huge.
/compact could lobotomize your agent and wasting 100K tokens actually hurt.
Now we have 1M context, cheap Opus 5.5 & Chinese models pushing prices toward zero.
So... who cares?
Just throw tokens at a problem - problem solved.
Except developers are doing the opposite.
Caveman just raised money & RTK is everywhere (even though it not really saving money).
So which future wins?
A) Tokens become basically free and optimization dies.
B) Jevons paradox: cheaper agents means 100x more usage, so efficiency matters even more.

ALT Line chart titled “Tokens Got 4× Cheaper. Spend Still Grew ~10×.” It shows OpenRouter token usage, spend and average price per token from Oct 2025 to Oct 2026, indexed to 1× in Oct 2025 on a log scale. Usage (green) climbs steadily to 39×. Spend (red) rises with usage to about 12× by July 2026, then dips to about 6× when the average price drops 75%, and recovers to 9.6× by October. Price (blue) stays near 1× until July 2026, then falls to 0.25×. Footer: a portrait engraving of economist William Stanley Jevons and the caption “Jevons Paradox: Cheaper tokens → more usage → efficiency still matters.” Source: OpenRouter via Token Elasticity (10.13.26).
1
31
Meta just declared the chatbot era over.
Muse Wonder is built to work for HOURS:
1M context
memory across conversations
delegates to sub-agents
calls tools in parallel
recovers from errors
keeps working through blockers
$4/$20 per 1M tokens.
Is the model war just shifted from
"who the best coding agent?" to "who can keep working without you?"
1
2
153
Be creative: you have a warehouse full of humanoid packed with company IP and you need to get rid of them. What do you do??
Figure AI asked the internet. Arnold Schwarzenegger said "You should melt them" So they trained the robots to jump into volcano 🌋
Finding a foundry willing to let robots with lithium batteries jump into its furnace took them all the way to Finland. And of course they filmed it, with Arnold on set because... well, why not?
Got a spare $2k? You can own a robot head made from the melted metal.
Honestly, this is one of the best ad stunts I’ve seen in a while. A (real?) reply from Arnold turned into a plan, robots trained in simulation to jump, six melts in 24 hours and a Terminator ending. I would have loved to be on that set
So, how would you have done it?
1
1
151
I could not make this shit up.
Hackers hid their crypto-mining army's control-center address inside a POEM on GitHub.
Victims pay the bills. Hackers keep the crypto.
Need to move the server? Edit the poem.
Clever hiding place, or a very dramatic address book?
1
41










