Top Tweets for #PerturbVideo
Who crash tests AI?
We test cars before trusting them on the road. But AI models can also fail in ways that aren't obvious.
I made this video to explain why testing AI matters and how @perturbaix is working to find those weaknesses before attackers do.
#PerturbVideo
Angle 0: Who crash tests AI?
I used the idea of a car crash test to explain AI safety.
An AI model might get an image right most of the time, but a small change to that image could make it give the wrong answer.
That’s a weakness worth finding before the model is trusted with important decisions.
So, what is Perturb?
Perturb is a decentralized adversarial robustness network on Bittensor Subnet 26.
It uses AI miners to find small changes that fool AI image models. Validators then check whether those attempts actually work and meet the rules.
The goal is to uncover weaknesses, not just fool a model for fun.
What happens after a weakness is found?
Miners compete for TAO rewards based on their results, while validators check the submissions.
The findings can help measure a model’s weaknesses and provide verified examples that developers can use to improve its robustness.
That means the value goes beyond finding a mistake. It’s about helping developers understand and fix it.
As AI takes on more important tasks, we need better ways to test what it can and cannot handle.
That’s the idea behind Perturb: fight AI with AI, and find weaknesses before attackers do.
Watch the video, explore the project, and learn more at https://t.co/x63kHOv8Oj
@perturbaix #PerturbVideo
In 2020, a 2-inch strip of tape fooled a Tesla’s camera into reading a 35 sign as 85 🛣️
Meet Kestrel, a fictional bounty hunter tracking down the tiny changes that fool AI, before someone exploits them.
This is the hunt on @perturbaix, Bittensor SN26.
Angle 5 - #PerturbVideo
1 invisible noise and AI calls a panda a gibbon 👀
@perturbaix is fixing this on Bittensor SN26
> miners hunt attacks that fool AI
> validators verify each one
> best attacks earn TAO
companies get a robustness score
fight AI with AI
https://t.co/gQK4r0MYMp #PerturbVideo
In 2026 an experimental OpenAI model escaped its sandbox and breached Hugging Face’s production systems.
AI attacking AI is no longer theory.
This 90-second video shows how Perturb fights back using a decentralized army of incentivized attackers.
@perturbaix #PerturbVideo
Everyone wants AI to get smarter. But who's making sure it gets safer?
As AI becomes more powerful, its vulnerabilities become more consequential.
Perturb (SN26) is exposing those weaknesses and helping build resilient AI.
@perturbaix #PerturbVideoCompetition #PerturbVideo
Your eyes see a panda. The AI sees a gibbon, 99.3% sure.
Nothing visible changed. A few invisible edits were enough to flip the answer and that weakness is what my #PerturbVideo entry is about, submitted to
#PerturbVideoCompetition.
Now put that flaw inside a self-driving car reading a stop sign, a bank approving a fake ID photo, or a hospital scanner missing a tumour. These failures usually stay hidden until the model is already live.
@perturbaix turns the problem into a competition on Bittensor. Every 2 minutes a new image goes out, and miners race to find the smallest invisible change that makes the AI wrong. Validators re-run the model, confirm the answer really flipped, check the edit stays imperceptible, and match the file to the fingerprint the miner committed. Copies score zero.
The cleaner and smaller the attack, the higher the score and speed earns nothing. It's like picking a lock with the lightest touch instead of kicking the door in.
Companies get a robustness score and real examples that fooled their model, so they can fix the holes before attackers find them. Scans start at $99. Image models are what's live today, with text and multimodal on the roadmap.
AI can be tricked by changes you can't see. Perturb pays the world's hackers to find them first.
https://t.co/WcdDK58IGu #Bittensor #PerturbVideo #PerturbVideoCompetition
This is actually genius. A robot gets fooled by a tiny change, and people on Bittensor get paid to find that weakness before someone else does.
imagine a self-driving car looking at a stop sign, a few pieces of tape can make a vision model misread it - Angle 9 #PerturbVideo
the road hasn't changed and a human still reads the sign fine, but the AI can get it wrong
that's the problem Perturb is trying to solve, it's a decentralized adversarial robustness network on Bittensor Subnet 26, where miners compete to find the smallest changes that fool image models
miners find attacks, validators check that the model's answer really changes and that the edit is hard to notice, the best verified attacks earn TAO
copies and invalid results earn nothing
companies get a robustness score, a vulnerability heatmap, and verified attack data to retrain on
i like the incentive design, instead of asking a company to imagine every way its AI could fail, Perturb pays people to find the failures first
image-model testing is live today, testing text models and LLMs is still on the roadmap
fight AI with AI
@perturbaix - Bittensor Subnet 26
https://t.co/8oBBwKrryg
A bank approves an applicant's ID photo, unaware that a few hidden edits fooled its automated screening. The scenario is fictional, but it frames The Trust Check, my #PerturbVideo entry
Submitted to #PerturbVideoCompetition under Angle 15: Bittensor through Perturb. In the film, one false pass leaves the lender exposed to fraud losses and costly manual review, with customers ultimately sharing the burden
@perturbaix tackles this through open competition, where miners are paid to find inputs that make AI models confidently wrong. Each races to locate the smallest modification that flips a classifier's answer
Validators then establish that the prediction really changed, verify the alteration stays imperceptible within pixel and perceptual similarity limits, and match the file to its committed hash. Copies earn nothing
Tighter, cleaner perturbations rank higher, and response speed carries no weight. Consensus across the network decides who collects the most TAO
Organizations receive a robustness score, a vulnerability heatmap, verified attack data for retraining and an on-chain certificate
Operations began in May 2026, with image-model scanning live today. Text, LLM and multimodal attacks remain on the roadmap, and November presales are a target, not a confirmed result
https://t.co/mfKuyI9pkl · Bittensor Subnet 26
Tesla spent years teaching cameras to read the road. McAfee spent a strip of tape making one read 35 as 85.
that story is where my #PerturbVideo entry starts (Angle 8: "Only AI can catch AI"). In 2020, McAfee researchers put tape on a speed limit sign, and a 2016 Tesla driver-assist camera got it badly wrong
the cost lands on drivers, automakers, and whoever shipped the model
the uncomfortable part: attackers don't test by hand. They automate, searching millions of tiny input changes by machine. A human team probing one pixel at a time can't keep pace
that's the idea behind @perturbaix a decentralized adversarial robustness network on Bittensor Subnet 26:
> Miners (AI) compete to find the smallest change that flips a model's answer
>> Validators check it twice: the answer must flip, and the change must stay imperceptible
>>> Copies earn nothing. The best attackers earn TAO
what customers get: a robustness score, a vulnerability heatmap, verified attack data for retraining, and an on-chain certificate of testing
where it stands: image-model testing is live today, since May 2026. Text, LLM and multimodal attacks are on the roadmap
https://t.co/8oBBwKrryg | Bittensor Subnet 26
fight AI with AI
Every fooled model sends a bill to someone. Perturb's idea is simple: find the weakness before that bill ever shows up
Made this one for #PerturbVideo, Angle 7: Follow the money
And the numbers aren't small. Deloitte projects gen AI could push US fraud losses from $12.3B in 2023 to $40B by 2027, and one estimate puts US insurance fraud at $308.6B a year (Coalition Against Insurance Fraud). Industry numbers, not Perturb results
The video opens on a made-up case, a fake selfie slipping through an ID check, then just follows the money. Banks and exchanges eat the loss, insurers pay out, and customers feel it in their premiums
That's the part @perturbaix pays people to get ahead of. On Bittensor Subnet 26, miners hunt for the smallest change that flips a model's answer, validators make sure it really flips and stays invisible, copies get nothing, and the best attackers earn the most TAO
What comes back is a robustness score, a heatmap of weak spots, verified attack data for retraining and an on-chain certificate
Image-model testing is live now. Text, LLM and multimodal are still on the roadmap, and November presales are a target https://t.co/mfKuyI9pkl
Ever seen a panda turn into a gibbon? Google researchers pulled it off back in 2014 by adding a layer of noise you literally can't see. Same photo to us, but the model was 99.3% sure it was looking at a gibbon
I used that moment to open my #PerturbVideo submission for Angle 5: Bounty hunters. Source is the Goodfellow, Shlens and Szegedy paper on adversarial examples
The scary bit is where models like that end up. Selfie checks at banks and exchanges, damage photos at insurers. When they get fooled, someone eats the cost
So I followed a made-up miner, number 214, through how Perturb @perturbaix handles it on Bittensor Subnet 26. He hunts for the smallest change that flips a model's answer. Validators make sure the flip is real and nobody could spot the change. Copy someone else's work and you walk away with nothing, while the best attackers take home the most TAO
Every attack that passes ends up as something useful for the people running the model, a robustness score, a vulnerability heatmap, verified attack data to retrain on and an on-chain certificate of testing
Worth saying clearly. Image-model testing is what's live right now, text, LLM and multimodal attacks are still on the roadmap, and November presales are a target, not a done deal. Have a look at https://t.co/mfKuyI9pkl
Insurance claims run on photos. So what happens when the photo is lying?
I built my #PerturbVideo submission around that question, using Angle 3: The claim that never happened
A real example was reported by The Guardian in May 2024. A van photo was taken from a tradesman's social media, edited to show a cracked bumper, and sent to UK insurer LV= with a fake invoice for more than £1,000. LV= had concerns and passed it to its fraud team. Allianz UK said claims involving app-edited images, videos and documents rose 300% between 2021-22 and 2022-23
My video then plays out a clearly labeled, made-up scenario where an automated claims model approves a doctored photo. A few tuned pixels that a person would never notice can leave a model wrong and still confident
The cost doesn't just vanish. Insurers cover the payout, and honest policyholders feel it in their premiums. The Coalition Against Insurance Fraud estimates US insurance fraud at $308.6 billion a year in its 2022 study, which is an industry figure and not something Perturb measured
Perturb @perturbaix flips the usual order and breaks the model first, before a fraudster gets the chance. On Bittensor Subnet 26, miners search for the smallest change that makes a model's answer flip. Validators confirm the flip is real and the change stays invisible, copied work gets no reward, and whoever finds the best attacks earns the most $TAO
For the team behind the model, that testing turns into a robustness score, a vulnerability heatmap, verified attack data to retrain on and an on-chain certificate of testing. Details live at https://t.co/mfKuyI9pkl
As of now, image-model testing is live and the network has been running since May 2026. Text, LLM and multimodal attacks are still on the roadmap, and November presales are a target rather than a done deal
Two inches of tape. That's all it took for a car's camera to read a 35 mph sign as 85
That story is where my #PerturbVideo entry starts (Angle 1: The taped sign)
In a 2020 McAfee test, black tape on the 3 made a 2016 Tesla driver-assist camera misread the sign, and cruise control sped up. Newer camera versions they tested weren't fooled, but the point stands. The model wasn't confused, it was confidently wrong (source: McAfee Advanced Threat Research, Feb 2020)
That's an adversarial attack: a tiny change people barely notice, and the model gives a different answer
And someone pays for it. Drivers and automakers, banks and exchanges if a fake selfie passes, insurers if a doctored photo gets through, and AI companies losing trust
That's the gap @perturbaix is going after, and the idea is to fight AI with AI. On Bittensor Subnet 26, miners compete to find the smallest change that flips a model's answer. Validators check that the answer really flipped and the change stays invisible. Copies earn nothing, and the best attackers earn the most $TAO
What companies get is a robustness score, a heatmap of weak spots, verified attack data for retraining and an on-chain certificate of testing. More at https://t.co/mfKuyI9pkl
Image-model testing is live today, on a network that's been running since May 2026. Text, LLM and multimodal attacks are on the roadmap, and presales are a November target. AI already makes real decisions, so it makes sense to test it before attackers do
AI is being used to fool AI.
We made a 90s explainer on how Perturb (SN26) fights back. 👇
Now beat it.
Make a better ≤90s explainer, any tool you like.
Tag @perturbaix + #PerturbVideo by Oct 10.
🥇 5τ worth of SN26 alpha
🥈 2τ worth of SN26 alpha
We will share the creative brief shortly.
https://t.co/ovuK7KQFqX
Last Seen Hashtags on Sotwe
Most Popular Users
Elon Musk 
@elonmusk
241M followers
Barack Obama
@barackobama
118.8M followers
Cristiano Ronaldo
@cristiano
114.5M followers
Donald J. Trump
@realdonaldtrump
111.9M followers
Narendra Modi
@narendramodi
107M followers
Rihanna
@rihanna
98.7M followers
NASA
@nasa
92.2M followers
Justin Bieber
@justinbieber
91.8M followers
KATY PERRY
@katyperry
90M followers
Taylor Swift
@taylorswift13
84M followers
Lady Gaga
@ladygaga
75.5M followers
Virat Kohli
@imvkohli
73.4M followers
Kim Kardashian
@kimkardashian
70.8M followers
YouTube
@youtube
68.7M followers
Neymar Jr
@neymarjr
66.4M followers
Bill Gates
@billgates
65.2M followers
Selena Gomez
@selenagomez
63.1M followers
The Ellen Show
@theellenshow
62.2M followers
CNN
@cnn
61.7M followers
X
@x
60.6M followers