Sagar Tanur ๐Ÿ‡ฎ๐Ÿ‡ณ (@Sagarvd01) header
Sagar Tanur ๐Ÿ‡ฎ๐Ÿ‡ณ (@Sagarvd01) profile picture

I build AI tools and break them. Notes on agent, MCP and dev tool security from the builder side. Making Layerize and SaaSecure.

Tanur
Joined November 2013
Photo by Sagar Tanur ๐Ÿ‡ฎ๐Ÿ‡ณ (@Sagarvd01)Photo by Sagar Tanur ๐Ÿ‡ฎ๐Ÿ‡ณ (@Sagarvd01)Photo by Sagar Tanur ๐Ÿ‡ฎ๐Ÿ‡ณ (@Sagarvd01)Photo by Sagar Tanur ๐Ÿ‡ฎ๐Ÿ‡ณ (@Sagarvd01)Photo by Sagar Tanur ๐Ÿ‡ฎ๐Ÿ‡ณ (@Sagarvd01)Photo by Sagar Tanur ๐Ÿ‡ฎ๐Ÿ‡ณ (@Sagarvd01)Photo by Sagar Tanur ๐Ÿ‡ฎ๐Ÿ‡ณ (@Sagarvd01)Photo by Sagar Tanur ๐Ÿ‡ฎ๐Ÿ‡ณ (@Sagarvd01)Photo by Sagar Tanur ๐Ÿ‡ฎ๐Ÿ‡ณ (@Sagarvd01)Photo by Sagar Tanur ๐Ÿ‡ฎ๐Ÿ‡ณ (@Sagarvd01)
CVE-2026-104120 is the official MCP fetch server (mcp-server-fetch) letting the model pick URLs that hit localhost, internal ranges and cloud metadata, and it follows redirects without re-checking where they land. the README warned about this before it had a CVE number. the draft fix PR got closed unmerged today and the latest PyPI release (2026.8.18) still has the old path, so upgrading does not save you. until a real fix ships, keep fetch away from agents that read untrusted pages, tickets or email, and block egress to 169.254.169.254 and private ranges.
1
7
896
a reset code is only safe if nobody can keep guessing it. Dooray, a Korean work suite, says an attacker kept entering password reset codes until one passed, then took over accounts and stole work files. rate limit your OTPs.
34
local model or API for your daily coding? one word answer.
32
what's the oldest API key that still works in one of your accounts right now?
19
i once set Layerize to 8 depth layers by default. looked great in screenshots, useless in real work. the middle layers were noise. now every layer has to be something you'd actually select.
14
fake depth of field in two layers. split this desk shot into near and far with a depth model, kept the front of the sketchbook sharp and blurred everything behind it.
25
coding agents are moving into CI, and JetBrains got one detail right. their experimental TeamCity plugin runs Codex or Claude Code inside a build, hands it provider credentials only for that build, removes them after, and saves the agent's full trajectory. short lived keys plus a transcript of what the agent did. that should be the default everywhere.
3
4
188
open weights doesn't always mean you can ship it. Edge0-145B says it runs on one 24 GB consumer card (streaming experts from system RAM). nice. but its license says selling the model, running a commercial model service, or offering a commercial coding assistant needs a separate written license from Edge0. read the LICENSE file before the benchmark table.
1
3
85
llama.cpp is 1.0 now, after about 3.5 years. the big change isn't a feature. from here on, breaking changes to the public C API come with a major version bump. boring news. boring is exactly what you want under every local AI app. what are you running on it?
2
1
48
if your app still calls claude-sonnet-4-5-20250929, it has 14 days. Anthropic retires it on Nov 30. after that, requests to it fail. their recommended replacement is claude-sonnet-5-5. search your repo, your env vars and your old cron jobs. the forgotten script is the one that breaks.
1
40
a firewall that runs anyone's command as root, as long as they ask through a file in /tmp. CVE-2026-123039: Check Point's Gaia OS has a root service listening on /tmp/dynamic_socket, writable by everyone, no check on who connects. attackers got in through the Mobile Access portal first, then used this to own the box. CISA added it to KEV on Nov 13. the fix is basically 0777 to 0660. permissions still matter.
1
65
you never signed up for Fireworks AI. your data might have been there anyway. Notion AI chats, Cursor agent traces, and now Uber drivers' license info all got hit through that one vendor.
39
one leaked vendor key ended with a fake Claude Code extension in Cursor's marketplace. Cursor's incident report: a key it gave Fireworks got exposed. old agent traces in a shared bucket had API keys nobody redacted. those keys opened more doors, and one let the attacker swap in a tampered extension that tried to plant an SSH backdoor. live about 11 hours, up to 5,000 users. a key pasted into an agent chat is still a live key.
1
86
be honest. have you ever read a SaaS company's subprocessor list before signing up?
15
i officially turned 31 today. before turning 18, i felt like time was slow af and wanted time to move faster because i want to reach 18 to get a driving license. now, time it flying and i wish it could be a bit slower here's what i looked like when i turned 18.
20
vendor breaches keep happening. this is why Layerize never sees your photos. the depth model runs in your browser. your image doesn't get uploaded to my server, so there's nothing for me or a vendor of mine to leak. trade off is a slower first load. i'll take that.
14
read-only should be the default for every agent tool. write access is something it earns.
14