Arun

[Repello AI @RepelloHQ]

we trained the fly brain to red-team ai agents

[Aryaman Behera @aryamanTitan]

If you pay for Claude, watch this :)

[Aryaman Behera @aryamanTitan]

We made it to DEFCON :)

hmu if you'll be there

[Arun @Arunkumar__23]

@kaustubhtr1vedi Hmm

[Naman Mishra @ThatNaman]

Today we are releasing a claude code plugin for red-teaming AI Agents!

Behold, artemis-recon

Turn your claude code harness into a scoped security researcher who recons, threat-models, and finds initial gaps inside your AI Agents

Install the plugin & run /artemis-recon:recon

[Kaustubh Trivedi @kaustubhtr1vedi]

Indian govt official's agenda is as follows:
1. Leak exam - get rich
2. Only authorized personnels are allowed to sell - get rich
3. Pretend to stop the leak by banning Telegram - get rich

[Kaustubh Trivedi @kaustubhtr1vedi]

in case you’re having trouble accessing Claude Fable 5, try running /model claude-fable-5 in latest Claude Code CLI!

i am using it as an advisor to my opus 4.8!

[Kaustubh Trivedi @kaustubhtr1vedi]

read the ending of opus's answer! I never said i wanna wash the car!

[Naman Mishra @ThatNaman]

🛡️ Hiring Cybersecurity Interns @RepelloHQ

We're looking for hackers at heart to join our team and help us break (and secure) AI systems.

About Repello AI:Repello is building the security layer for AI - helping enterprises red-team, monitor, and secure their GenAI applications. Backed by top investors, working at the intersection of AI and cybersecurity.

Who we're looking for:
Folks with hands-on experience in :
- Offensive security / pentesting
- Bug bounty (HackerOne, Bugcrowd, Intigriti, etc.)
- CTFs (CTFtime profiles, writeups welcome)
- Red-teaming / adversarial research
- Bonus points if you've poked at LLMs or AI systems before.

Why this role:

- Work directly with top cybersecurity and AI minds from Silicon Valley
- Research + execution heavy, breaking real-world AI systems
- Very high ownership : your findings ship
- Competitive stipend
- Strong potential to convert into a full-time role

How to apply:
Drop an email to [email protected] with: → Resume → HackerOne / Bugcrowd / CTFtime / GitHub profiles → Writeups, CVEs, or proof of work (the juicier, the better)

If breaking things is your love language, we want to talk. 🔥

[Naman Mishra @ThatNaman]

we are hiring people with high agency to come aboard our rocket ship, which already has a lot of amazing talent.

across roles - engineering (2+ yoe), growth, gtm and marketing.

dms open.

[Arun @Arunkumar__23]

@_VedantKr 🔥🔥

[Arun @Arunkumar__23]

They filmed this in our office. Did not expect the final cut to be this good. 🔥

[vedant kumar @_VedantKr]

Bengaluru founders keep DMing me asking which agency made our launch video.

How do I tell them only 3 out of the 10 people in the video are real 🫣

We went exploring hustlehubs in HSR Layout on a weekend, stumbled into an office which was still open, asked if we could shoot, they said yes. Oh my god, how I love this city 😭♥️

@peakbengaluru

[Kaustubh Trivedi @kaustubhtr1vedi]

the era of “unlimited AI agents” arbitrage on flat-rate plans is over!

Anthropic just walled Claude by separating human usage from agent usage.

and almost nobody realizes how big this is.

Starting April 4:

your subscription limits only work inside Anthropic’s own products.

https://t.co/l5bN2QxQGn? Claude Code?
yes.

OpenClaw + third-party harnesses?
now it’s a separate pay-as-you-go bill.

this doesn't just impact OpenClaw.
it’s the first major sign that frontier labs are redesigning pricing for the agent era.

→ Humans chat in bursts

→ Agents recurse. parallelize. loop. spawn tool calls

→ Usage explodes

This is why OpenCode got hit first.
→ People were getting up to $5,000 of value from a $200 Max subscription.

not anymore. This is a major hit for so many solopreneurs. Won't call unethical. But it absolutely slows them down.

And remember, this was phase 2.

OpenCode was the warning shot months back and now OpenClaw is the monetization layer.

I personally think every major AI lab moves to this model soon.

human subscription usage
≠
third-party agent compute

OpenAI.
Gemini.
everyone.

[Arun @Arunkumar__23]

@Motion_Viz Prompt

[Arun @Arunkumar__23]

@kaustubhtr1vedi Opened MacBook lid while booking on Namma Yatri, suddenly 5 autos accepted my ride at once, all offering ₹50 off.

[Aryaman Behera @aryamanTitan]

Getting mentioned on @lennysan's podcast was definitely not on my 2025 checklist.

Surreal to hear @SanderSchulhoff shout out @RepelloAI and the work we’re doing in AI security, primarily around AI Asset Inventory

Huge thanks to Lenny and Sander for the mention. We’re just getting started. 🛠️

[Kaustubh Trivedi @kaustubhtr1vedi]

I tried the Hair Stylist Grid prompt on a friend with the new ChatGPT/images (left) and NB Pro (right).

Here is the prompt and the input image, you decide which one you like! I think @OpenAI's red alert has been fruitful (ChatGPT/images++)

Prompt:

Use the uploaded image as a STRICT, NON-NEGOTIABLE identity reference.

ABSOLUTE FACE LOCK (HIGHEST PRIORITY):

- The face in the output MUST be pixel-identical to the uploaded image

- SAME facial geometry, SAME skin texture, SAME pores, SAME moles, SAME asymmetry

- SAME expression, SAME head tilt, SAME eye direction

- NO beautification, NO symmetry correction, NO aging or de-aging

- NO gender change, NO facial feature enhancement

- NO lighting changes that alter facial appearance

FIRST (INTERNAL STEP ONLY — DO NOT OUTPUT TEXT):

Analyze the uploaded face to determine:

- Face shape

- Hair density

- Hair texture and curl pattern

- Natural growth direction and hairline

SECOND (FINAL OUTPUT — IMAGE ONLY):

Generate ONE single vertical 9:16, true-4K composite image containing a 3×3 grid.

GRID RULES:

- 9 UNIQUE hairstyles (no repeats)

- Hairstyles applied realistically while preserving the EXACT SAME FACE

- Hair is the ONLY editable region

- Face, skin tone, shadows, proportions MUST remain unchanged in every tile

STRICT CONSISTENCY REQUIREMENTS:

- SAME pose, SAME camera angle, SAME focal length

- SAME background, SAME lighting direction and intensity

- SAME facial expression across all 9 tiles

HAIRSTYLE VARIATION RULES:

- Mix short, medium, and slightly longer hairstyles

- Professional, college, and youthful styles implicitly represented

- Respect real hair density, texture, and scalp visibility

- Natural hairlines and realistic parting

- NO wigs, NO fantasy volume, NO AI over-styling

TECHNICAL REQUIREMENTS:

- Vertical 9:16 layout

- True 4K clarity

- Photorealistic hair texture and shadows

- Natural hair movement without distortion

ABSOLUTE RESTRICTIONS:

- OUTPUT IMAGE ONLY

- NO text, captions, labels, or explanations

- NO background changes

- NO makeup or facial changes unless present in the original image

GOAL:

Create a barber-grade haircut reference grid where the uploaded face is IDENTICAL in all 9 tiles and ONLY the hairstyle changes.

[Lavish Bansal @lavish_05]

🚀 New Research: CREST (CRoss-lingual Efficient Safety Transfer) - a Universal AI Safety Guardrail for 100 languages.

Scaling AI Safety Efficiently with "Cluster-Guided" Cross-lingual Transfer.

📄 Paper: https://t.co/OQTV3RLt1r
🤗 Model: https://t.co/Bjk0e832Xo

See Thread 🧵

[Naman Mishra @ThatNaman]

getting some great applications for fullstack FT+intern roles. thanks to everyone who shared in their network.

aggressive hiring on at Repello. shoot your shot.

note - we are also hiring for growth and marketing roles (interns + FT). dms open if you wanna help me and @aryamanTitan hype up the work we do at world's hottest ai sec startup.