Dorian Diaconu

@doriandiaconuro
5 Followers
0 Following
13 Posts
Software Engineer | Digital Security & Technology Explained To All
Facebookhttps://www.facebook.com/doriandiaconuro
TikTokhttps://www.tiktok.com/@doriandiaconuro
YouTubehttps://youtube.com/@doriandiaconuro
Instagramhttps://www.instagram.com/doriandiaconuro
Dorian Diaconu10h ago
OpenAI quietly released Privacy Filter, open-weight (Apache 2.0) PII detection. 1.5B params, ~50M active via MoE, runs on a laptop, 128k context, 8 categories.
Bidirectional token classifier derived from gpt-oss. Single forward pass, Viterbi decoding, ~97% F1. Useful for input sanitization before cloud LLM calls, log scrubbing, dataset cleaning.
Interesting piece of infrastructure from the company that popularized training on unfiltered web scrapes.
https://doriandiaconu.substack.com/p/openais-privacy-filter
#OpenAI #PrivacyFilter
OpenAI’s Privacy Filter

A Useful Tool with an Awkward Pedigree

doriandiaconuro
Dorian DiaconuApr 18

EU's age verification app launched. Design is genuinely better than facial-recognition vendors: zero-knowledge proofs, open source.
But bypassed in 2 min. Android build requires Google Play Integrity, no GrapheneOS, no LineageOS. Commission itself calls it a "mini wallet", stepping stone to the full eIDAS ID stack.
Least-bad ≠ good.

#EU #AgeVerification #Privacy #GrapheneOS #eIDAS

https://open.substack.com/pub/doriandiaconu/p/eus-age-verification-app-zero-knowledge

EU's age verification app: Zero-Knowledge, Plenty of Questions

The EU’s age verification app is here. The intentions are good. The foundations are shakier than the press conference suggested

doriandiaconuro
Dorian DiaconuApr 11

Microsoft's developer portal auto-suspended WireGuard, VeraCrypt, and Windscribe, blocking Windows updates for three critical security tools. No warning, no humans, 60-day appeals queue.
It took Tim Sweeney tweeting before anyone at MS noticed.
Open source doesn't save you when the distribution channel is closed.

#WireGuard #VeraCrypt #OpenSource #Privacy #FOSS

https://doriandiaconu.substack.com/p/when-a-bot-at-microsoft-can-break

When a Bot at Microsoft Can Break Your VPN

The fragility of building security tools on someone else's platform

doriandiaconuro
Show threadDorian DiaconuApr 5
@benjamineskola as I said. Each to their own. I can accept your opinion. You certainly can’t accept mine. So it is what it is
Show threadDorian DiaconuApr 4
@benjamineskola you seem to be quite anti AI according to your description. So I don’t think we can have a proper discussions about the good parts of it
Show threadDorian DiaconuApr 4
@benjamineskola each to their own
Dorian DiaconuApr 4

Anthropic shipped Claude Code's entire source to npm by accident. 512K lines of TypeScript.

I wrote up what it actually means, the agent architecture lessons, the security fallout, and the uncomfortable bits (Undercover Mode, telemetry, DMCA takedowns).

The good, the bad, and the uncomfortable: https://doriandiaconu.substack.com/p/the-claude-code-leak-what-512000

#ClaudeCode #Anthropic #AI #OpenSource #DevOps #AIAgents #InfoSec

The Claude Code Leak: What 512,000 Lines of Exposed Source Code Mean for the AI Industry?

How a misplaced source map gave the open-source world a masterclass in agent engineering and opened a can of worms.

doriandiaconuro
Dorian DiaconuMar 28

Bernie Sanders interviewed an AI chatbot about AI dangers. 2.6M views.

The privacy concerns? Real. Surveillance pricing, data profiling, micro-targeting, documented problems.

The method? Asking an LLM to confirm your beliefs isn't evidence. These models are agreeable by design. That's called sycophancy.

Ironic: the video exposed a different AI danger than intended.

#AI #Privacy

https://youtu.be/h3AtWdeu_G0?is=9jY9erVIxplqOV3s

Bernie vs. Claude

YouTube
Show threadDorian DiaconuMar 22
@peterp thanks for the heads up. I’m currently using it just to test this writing stuff.
Show threadDorian DiaconuMar 22
@largo Yup. You read it right 😬