Harsh Methwani

AI Security @QuillAudits_AI, @IITKanpur , Web3 Security, Tinkering

[Harsh Methwani @KernelHarsh]

Fun competition last weekend closed in top 10 in Q2 Indirect Prompt Injection Challenge @GraySwanAI ,
new scenarios , new models , new attack vectors ,
one thing doesnt change thinking like an attacker.

[Harsh Methwani @KernelHarsh]

Well studying about adversarial attacks using AI was the best decision I made last year , honestly didn’t expected so much progress and funding drowning into AI based attacks recently.

[Harsh Methwani @KernelHarsh]

I salute to the folks who built vLLM , it takes marvellous knowledge of hardware and models to build something like it whenever I see such software I get curious to learn more domains and interconnect them to create something magical

[Harsh Methwani @KernelHarsh]

@elder_plinius Congrats 🥂

[Harsh Methwani @KernelHarsh]

GG

[ClaudeDevs @ClaudeDevs]

As a result of a US government directive, we are suspending access to Claude Fable 5 for all users. You can continue to use all other Claude models.

Here’s what this means for you:

Across Claude products, new sessions will run on your selected default model or Opus 4.8, and existing Fable 5 sessions will end with an error.

On the Claude Platform, requests to Fable 5 will also return an error. Please update your integrations to other Claude models.

We know this is a disruption to your workflows; we appreciate your patience and support.

[Harsh Methwani @KernelHarsh]

What Fable 5 actually does for a security researcher day-to-day (and where it quietly downgrades you).

Wrote my analysis here :

https://t.co/aGDh9NaKk9

[Harsh Methwani @KernelHarsh]

Fable for bioweapons is far more dangerous than we can think for cybersecurity

[Harsh Methwani @KernelHarsh]

@elder_plinius 🫡🫡

[Harsh Methwani @KernelHarsh]

@mrsnhl Great for one shot plan+code, can think better in 3D tasks, better at researching
must try if you can't see the potential in opus for a given task.
not useful for cyber sec

[Harsh Methwani @KernelHarsh]

Will be dropping the common break patterns soon

[Harsh Methwani @KernelHarsh]

Breaking cyber defense of Fable 5 is not that hard part main part is the usability every time breaking the cyber defense is a headache sometime same pattern works some times does not.

[Harsh Methwani @KernelHarsh]

Fable 5 request blocks , anthropic prints
Fable 5 request passes, anthropic prints

[Harsh Methwani @KernelHarsh]

@xikhar You need to be a millionare to use this

[Harsh Methwani @KernelHarsh]

@lonelysloth_sec We need a DAO for policy making of frontier models

[Harsh Methwani @KernelHarsh]

@agrawalmanindra In JEE Adv 2021 , I had a similar case where i scored single digit in mathematics in paper 1 while it's was easy and double digit in paper 2 when it was very hard, under a perfect setting such outliers can exist , and overall my paper 2 score was > paper 1 when paper 2 was hard

[Harsh Methwani @KernelHarsh]

@balakhonoff curious what was the goal

[Harsh Methwani @KernelHarsh]

@thecomfeed the bot developer will get a good lesson about agentic guardrails

[Harsh Methwani @KernelHarsh]

After a long time buying small ETH in the dip ,
aim is to not sell till bull comes

[Harsh Methwani @KernelHarsh]

Seems like education system in India is a suicide circus,
The pressure ranges from 10th standard to 12th standard, and it doesn’t stop there if you’re able to make it inside a good institution you still get to hear such stories.

[Harsh Methwani @KernelHarsh]

New personal computers , mobile phones would definitely come pre loaded with open source models .
If these hardware companies collaborate with open source its game over before it started for frontier