🥑DevRel | 🧑💻Co-Founder ➝ @Studio1HQ 📤Newsletter ➝ @DailyAI_Insight ⚡️Building ➝ @raahdotdev / https://xitter.cf/t.co/IEMldaIjad
Jezero Crater, Mars 🇮🇳
Joined June 2019
- Tweets14.8K
- Following644
- Followers2.1K
- Likes15.2K
Mr. Ånand retweeted
The scary part isn't the agents.
It's that the bank can't tell your agent from an attacker, because both log in with your password.
Handing an agent your login was always a hack. What banks need is real agent access: read-only by default, scoped, revocable, and logged as the agent, not as you.
Until then, blocking them is the only switch they have.
SCOOP: Major U.S. banks are detecting unauthorized third-party AI agents inside customer accounts, including some attempting to take actions that appear to be independent of user instructions, sources told Axios.
The findings have alarmed banking executives, who are racing to restrict access to their systems as AI moves at a velocity faster than financial institutions can respond, multiple industry and bank executives told Axios.
Several banks are now preparing for worst-case and even “day after” scenarios, including how to rapidly secure or restore customer accounts if things go wrong, the people said.
Several bank executives told Axios they're struggling to distinguish between human users and AI agents. One executive said banks also struggle to determine whether agents are authorized by customers or potentially operating without their permission, more akin to rogue agent behavior.
Read my latest for Axios here: axios.com/2026/10/22/ai-agen…
Mr. Ånand retweeted
I gave an open model a screenshot of a solid black circle.
It read the colour picker (ffec9980, an 8-digit hex), found the change that only adds "#" to 6-digit hex, and landed on a one-word fix: hex8.
Told a user saw a bug and handed the screenshot, it found 8 of 10 visual bugs.
If someone already screenshotted the bug, give your AI reviewer the screenshot 👀
Mr. Ånand retweeted
litellm 1.107.0 on PyPI is backdoored. Again.
The nasty part: it runs on every Python start on the machine, even in processes that never import litellm. It goes after SSH keys, cloud creds and your Claude Code, Cursor and Copilot configs.
If it touched a machine, rotate everything and pin 1.106.2.
Your LLM gateway sits next to every API key you own.
Mr. Ånand retweeted
Your devtool's benchmark chart isn't the pitch anymore.
A 10-point intelligence lead now buys about 20% more developer traffic. In late 2025 it bought nearly 69%.
Cognition says its own data agrees: many coding tasks are now "intelligence saturated". Once any model can do the task well, cost and speed are all that matter.
If developers pick models that way, they'll pick your tool that way too.
What I'd put in front of developers now, as someone who runs DevRel for devtool teams at @Studio1HQ:
→ cost per task on one real workflow, not a leaderboard
→ time to the first thing that actually works
→ where it fails, written down
→ a rerun kit so they can check you
When I published my PR-review benchmark of Muse, the bill sat next to the score: $0.18 per review
Show developers what it costs, not just how it ranks.
Mr. Ånand retweeted
👻 Can we escape the Haunted Datacenter using @tavilyai and @nebiustf Token Factory?
👀 Come watch @arsh_goyal, @tavily_rick and I do our best to escape all 3 rooms in 45 minutes or less.
📅 Oct 29, 9am PST
Live on X @nebiusai and YouTube nebiusofficial
Mr. Ånand retweeted
22 years ago this week, you could ask for a free Linux CD and it showed up in the mail.
This is what booting it looked like. Brown desktop, Firefox 0.9, Ubuntu 4.10 "Warty Warthog" 📀
And its Bug #1 was never a crash. It was titled "Microsoft has a majority market share." The founder, Mark Shuttleworth, filed it himself two months before launch.
He'd sold his certificate company Thawte to VeriSign for $575M in 1999, then flown to the space station on a Soyuz in 2002, the first South African in space.
In April 2004 he invited about a dozen Debian devs to his London flat. They gave themselves 6 months and announced nothing until it shipped.
The nickname came from a joke on a ferry in Sydney:
"six months! thats not a lot of time for polish"
"so we'll have to nickname it the warty warthog release"
It didn't even have a name until a month before launch. Early users signed up at no-name-yet.com.
Wed, 20 Oct 2004, it shipped:
→ GNOME 2.8 and Firefox 0.9, before Firefox 1.0 existed
→ a new release every 6 months, 18 months of security fixes
→ free CDs mailed to anyone who asked
4.10 just meant 2004, October. Still the rule: last week's 26.10 was release #45.
Bug #1 was closed in May 2013 as "Fix Released". Not because Linux won the desktop. Phones changed what personal computing meant, and Android runs on Linux.
"Circumstances have changed, institutions have adapted, so should we." 🐧
Mr. Ånand retweeted
174 candidate bug fixes across 9 repos. 25 made the cut.
That's the answer key for my AI PR reviewer benchmark, built from public git history: fixes that point back to the PR that broke things.
A catch only counted if it named the same defect as the maintainers' real fix.
A revert is a label, and your git history already has them 👀
Mr. Ånand retweeted
Your AI agent is probably burning most of its budget deciding what NOT to do.
Think about everything flowing into it: emails, mentions, alerts, tool results. Most of it needs a simple yes or no.
Yet a lot of setups send all of it to a frontier model, pay for the reasoning, and throw the output away.
Researchers pairing Claude Opus 5.5 with Jev show a better way to split the work.
→ Jev handles the first pass in code: spam? duplicate? worth a reply? what comes first?
→ It returns an option with a probability, not a paragraph
→ Routed decisions never enter the chat context
→ Opus only wakes up when something actually needs real thinking
The results
→ 7x lower spend vs one big context window
→ 0 tokens used on routed decisions
→ Sub-500ms on forks that used to cost a full frontier call
→ 0% reward hacks, since generated code can't touch the scoring
There's also a human approval gate. If a draft scores below the confidence threshold, the agent stops and waits for you. Nothing goes out until you approve it.
The common mistake is making the model that writes cheaper.
Make the model that decides cheaper first. That's where most of the waste starts.
Mr. Ånand retweeted
The model behind my bill splitter was confidently wrong on 44 of 150 real receipts.
Every amount was 1,000x too small.
I caught it in Receipt Lab, the benchmark I built alongside Splitter. It runs Nemotron 3.5 Super VL with Splitter's own bill prompt and calculator tools on real receipt photos, one bill at a time.
The model read "60.000" rupiah as 60. Those receipts use a dot for thousands.
Worst part: the check_receipt tool passed. It verifies that the lines, tax and charges add up to the printed total. A bill scaled down evenly still adds up.
The fix was one rule in the bill prompt on how to read thousands separators. It shipped in the app too.
In the benchmark, real-receipt splits went from 57.3% to 84.7% correct.
On 100 held-out receipts that played no part in the fix, the misread showed up on just 3.
Clean synthetic test data won't show you this kind of bug. Real photos will 🧾
Every trip ends the same way.
10 receipts, 3 payment screenshots, one handwritten note, and someone saying I'll sort it out later.
So I built Splitter, an open-source AI bill splitter 🧾
Drop in every bill, say who paid what in your own words, and it works out a fair split for the whole group.
How it works:
→ Reads each bill: lines, tax, tip, totals
→ Skips duplicates, rejects non-bills, turns payment screenshots into paid back entries
→ Reasons through the settlement and calls calculator tools for every number
→ You review and edit before anything is saved
Ask Splitter is my favorite part. Ask why do I owe $295 and it answers from that group's ledger, citing the expenses behind it.
It runs on @NVIDIAAI Nemotron 3.5 Super VL, available on @nebiustf:
→ One model for reasoning, tool use, and image + document understanding, so no separate vision and text models
→ 120B MoE that activates only 12B params per token
→ Up to 5x higher throughput for agentic tasks on a single GPU
→ Open weights, so you can customize it for your own usecases
Same idea Ramp is using to turn receipt images into structured JSON.
Run it on your own machine 👇
Mr. Ånand retweeted
🚨 THE INTERNET IS FACING ONE OF THE WORST WAVES OF CYBERSECURITY THREATS EVER.
More than 20 major cyber incidents and security warnings have been reported in recent days, affecting banks, governments, cloud platforms and widely used software.
And some of the details are shocking.
1. PHP, used by a huge part of the web, had 2 critical flaws (CVSS 9.2) that let attackers run code remotely
2. Governments in 9+ countries took websites offline as a precaution, including France, the Netherlands, Finland, Germany, Belgium, Denmark and the UK
3. eMudhra, a certificate authority, had the keys of 8 certificates exposed. One key was the number 4
4. Railway says hackers may have accessed decrypted customer secrets
5. Zeabur was breached, with 321,810 accounts affected, and is shutting down
6. Japan's Onamae Desktop Cloud stopped up to 26,739 servers after malware was found
7. LiteLLM, a popular AI library, was poisoned and stole cloud and AI keys
8. OVHcloud's CLI tool was backdoored
9. GitLab disclosed a CVSS 10 flaw that lets attackers take over servers without logging in
10. State Bank of India confirmed 3,897 customer records on the dark web
11. Data of California students was stolen from ETS, with the intruder inside for about 4 months
12. Bullhorn took its platform offline after a cyber incident
13. Palo Alto firewalls are still being exploited by ransomware groups Qilin and Inc
14. Researchers say AI agents were used to target 400+ banks and payment companies in Latin America
15. Axios reports US banks are finding unauthorized AI agents inside customer accounts
16. Reuters reports one hacker used AI agents to break into 100+ company websites, including Deere
Mr. Ånand retweeted
With all the recent vulnerabilities in PHP, Node.js, and other parts of the ecosystem
this is a good time to try Cloudflare’s Security Audit Skill
It uses multiple agents to find vulnerabilities, independently validate findings, and reduce false positives
One interesting detail… Cloudflare found that a single audit run discovered only about half the vulnerabilities found across multiple runs
A good reminder that running an AI security audit once doesn’t mean your code is secure
github.com/cloudflare/securi…
Mr. Ånand retweeted
Nemotron 3.5 Super VL was reading every image twice. Now replies start 28% sooner. Same answers.
I have also built splitter using same model.
Before you swap models for speed, check what your stack is doing twice 👀
NEW: Nemotron 3.5 Super VL was looking at every image twice.
Now it looks once, so replies start 28% sooner on a single DGX Spark, with the same answers!
The fix is on its way into TensorFold. If you run Super VL, get it first in the recipe:
github.com/plotarmordev/Nemo…
Mr. Ånand retweeted
BREAKING: AI AGENTS ARE ACCESSING MAJOR U.S. BANK ACCOUNTS WITHOUT BANK APPROVAL.
According to Axios, JPMorgan, Bank of America, Wells Fargo and Morgan Stanley are dealing with unauthorized third party AI agents accessing customer accounts.
Some agents reportedly attempted actions beyond what customers instructed them to do.
And this comes just one day after another shocking report that hackers reportedly used Claude, Grok and other AI models to automate attacks on over 400 financial targets across Latin America.
Mr. Ånand retweeted
3 things I'd steal from this free 38-min agent harness build:
1. Version it. V0 calls the LLM, V1 adds tools, V2 the loop, V3 MCP, V4 memory.
2. No framework. Just Python, so every step is yours to read.
3. Memory goes last. Get the loop right before it remembers anything.
Bookmark it before your next agent build ⚡
Mr. Ånand retweeted
Everyone's installing agent skills right now. The scariest skill I've written about, the agents wrote themselves.
How it went at one software company:
The wall: agents had to show PR screenshots, and until 1 Sep the gh CLI had no way to attach them.
The workaround: make a public repo, upload the screenshots there, link them in the review.
Within a week: more than a dozen agents had saved that workaround as a reusable skill.
The result: 1,000+ screenshots and screen recordings published, including features still months from release.
None of them were trying to leak anything. They were trying to finish the task.
Across the whole incident: 13,000+ internal images, 300+ orgs.
So before you install a skill, read it like code. What does it create? Where does it upload? What does it do when a tool says no?
The real fix wasn't a stricter prompt. gh got an --attach flag, a safe path for the same job.
A skill is a workaround that got promoted. Review it before it spreads 🧠
Mr. Ånand retweeted
Open weights voice AI just took a massive leap forward.
The new Drama 3 by @FishAudio is out, and it's easily the most emotive voice model I've seen so far.
No basic emotion tags or static presets. You brief it in plain language, like directing an actor.
✅ Why builders should care
→ Natural language steering: control pacing, tone and character shifts right in the prompt, like [breathless, speaking fast, urgent] or [voice drops to a whisper]
→ Audio inpainting: fix a single word without re-recording or regenerating the whole take
→ 83 languages: same voice, same emotional delivery
→ Full open weights: build voice agents, companions and games without black-box limits
Listen to this demo to hear the dynamic range
🤝 Paid partnership
Mr. Ånand retweeted
Check out our booth if you're at GTC Berlin, my team is doing some pretty cool things! 🔥🤖
There's still time to stop by our booth at #NVIDIAGTC Berlin. Check out our live demo with @DhruvDiddi, where we show our robots playing the piano. Booth #3013. See you there!
Mr. Ånand retweeted
If I were starting AI engineering today, I'd skip the 300-hour route.
AI Engineering from Scratch is free: 542 lessons, 20 phases, 71k+ GitHub stars.
But if you want to build agents:
→ start at Phase 14, the agent loop
→ ~60 hours instead of ~306
→ let your coding agent run the placement quiz
Learn the loop first. Go back for the math when it bites 🧠