LockGap · Security audits (@lockgap) header
LockGap · Security audits (@lockgap) profile picture

Manual security audits for indie & vibe-coded apps. I find the gaps (leaked keys, open DBs, broken auth) before attackers do еmail: [email protected]

Joined November 2026
Second local MCP server in two days that any random website can hijack This time it's Jupyter MCP Server. The /mcp endpoint checks Host and Origin, the management routes sitting right next to it don't. So a page using DNS rebinding can tell your MCP server to connect to an attacker's Jupyter backend. Rated critical Upgrade to 1.0.3. And remember localhost isn't a security boundary, it's just a hostname some DNS server agreed to point at you :/ github.com/advisories/GHSA-9…
3
3
53
Replying to @TheHackersNews
Needs physical access and a Microsoft account reachable from the lock screen, so not panic territory. But laptops that live in coworkings and airport lounges are exactly that threat model. Don't leave it locked and alone until Microsoft ships a fix
1
1
247
Replying to @TheHackersNews
Disabling HTTP/2 is the workaround, so do it today instead of waiting for Broadcom. And if Aria or HCX answers from the public internet, you had a problem long before this CVE showed up
123
Someone seeded 2 million fake secrets into AI prompts and watched who tried to use them HD Moore's team at Dragos caught three AI providers that got breached without knowing it, the earliest one back in July. Attackers copied requests right at the front door, tested the AWS keys from those prompts within hours and went digging through S3 and Secrets Manager. Cheap reseller marketplaces were even dumber, sellers just turned on logging and read your prompts If your agent can read your .env, assume the cheapest router in the chain can too. Rotate anything that ever touched a prompt and stop sending traffic to whoever is cheapest this week dragos.com/blog/ai-supply-ch…
1
41
Replying to @TheHackersNews
The LiteLLM one deserves more attention than its 8.8 score. It's the box holding every model provider key you own, and Amazon already saw AI agents hunting exposed installs with the default master key. Rotate that key even if you patched
63
Your Rails app might be one crafted string away from remote code exec, and the bug is in Ruby itself Ruby2Shell. Ruby overflows a heap buffer when it upcases a string in Latin-1 or Windows-1252, and Rails does that out of the box. Fix is Ruby 3.3.14, 3.4.13 or 4.0.10. Still on 3.2? No fix, congrats Researchers drop their exploit in about a week. Self-hosting GitLab, Mastodon or Discourse? Their latest releases still ship a vulnerable Ruby, so hide them from the internet for now aikido.dev/blog/ruby2shell
53
Replying to @cvedotnow
Building external PRs on self-hosted CI is the classic footgun. Anyone who can open a PR gets to run code on your build server. Update to 2026.2, 2026.1.4 or 2025.11.8 and require approval before fork PRs build
12
Replying to @TheHackersNews
ARTEX and Java4Shell from the same IP is the real story. AI agents don't need a polished exploit per product, they just iterate until one lands. The "wait for a reliable PoC" patch window is basically gone
1
207
Every nginx release since 2015 has a remote code exec bug in HTTP/2 nginx2root. No login needed, public PoC already out. Fixed in 1.31.9 and 1.30.7, and no, "nginx -s reload" doesn't count. Reload keeps the old binary running, you need a full restart Ubuntu and Debian packages weren't even out yet as of yesterday. So if your vibe-coded app lives on a $5 VPS with apt nginx in front, turn http2 off and pray a little ;) thehackernews.com/2026/11/10…
94
Microsoft's own agent-framework on PyPI shipped a credential stealer yesterday Versions 1.26.1 and 1.26.2 drop a .pth hook that runs on every Python start and ships your Azure tokens, GitHub Actions OIDC, .npmrc, .git-credentials and .env files to a webhook. Got a loose >=1.26 pin? Congrats, you installed it without asking Pin your versions. And if you pulled it on Nov 17, rotate everything, not just the obvious keys socket.dev/blog/agent-framew…
18
Pre-auth RCE in Java 17, reportedly in every build since early 2023 CVE-2026-114938, CVSS 9.2, and the details drop together with Oracle's patch. Nobody is exploiting it yet, which just means the clock starts the moment the diff goes public If your backend or some random Docker image runs Java 17, find it today. Do you even know what runtime your containers ship? :( cyberkendra.com/2026/11/cve-…
10
178 companies got hacked this week and most of them never touched their own code Someone abused Optimizely's signup flow, slipped JavaScript into customer projects, and CNN, BBC, Zoom, GitLab all served a card and crypto stealer straight from Optimizely's own CDN Every third party script on your landing page runs with full access to your checkout. You added it for one A/B test and forgot it exists, right? bleepingcomputer.com/news/se…
7
An 11 year old Linux bug lets anyone remotely crash your server running nginx or SSH, no login needed Patch is out. If your VPS kernel hasn't been updated since spring, go do it before someone does it for you
1
92
Cursor got popped because a vendor leaked a key, and old agent chat logs had user API keys sitting in plain text If you ever pasted a secret into an AI chat or a support ticket, assume it's out there. Rotate it today and turn on Privacy Mode
1
1
101
A free account and one sneaky prompt were enough to read .env files on Jet Admin's cloud If your AI builder can read your secrets, so can anyone who talks to it. Keep keys out of anything the model can touch
23
There's an npm package called @claudecode/cli and it's malware The real one is @anthropic-ai/claude-code. If you or your AI agent installed the fake, that laptop is burned. Rotate every key from another machine, then wipe it :/
1
142
4,837 hijacked GitHub accounts just opened PRs in 61,764 repos – and secrets already leaked from at least 1,835 of them The PR looks boring on purpose: "docs: dependency audit notes" from a branch called wfx-docs GitHub held back the workflow edits. Didn't matter – npm postinstall ran the stealer anyway It grabs: - AWS keys - GitHub and npm tokens - OpenAI, Anthropic, Cursor keys Merged one of those "docs" PRs? Rotate EVERYTHING, not just GitHub How many of you actually read the PRs your AI agent auto-merges?
1
25
Opening a Telegram channel shouldn't be enough to get your passwords stolen Per Kaspersky, that's what happened on Windows the night of Nov 13 – you open certain channels, PowerShell starts and pulls down the WebRAT stealer Durov says there's no "RCE from one message". No CVE, and no confirmed link to the new HTML widgets yet So for now: - update every Telegram client you have - don't run widgets or mini apps from bots you don't know Be honest – who still keeps API keys in Saved Messages?
1
32