Hermes like the Agent not the handbag. Building Voice-AI Agents, Director of DevRel @Agoraio | Polyglot Engineer / Evangelist | Thoughts & opinions are my own
New York, USA
Joined June 2009
- Tweets10.3K
- Following994
- Followers1.8K
- Likes26.2K
Hermes ᯅ retweeted
An OpenSSH bug that took 46 years to ripen!
1980
UNIX puts * in the password field of system accounts. The idea is simple: it’s not a real hash, so it can’t be brute-forced and you can’t log in with a password.
2012
The musl libc developers decide that returning NULL from crypt() on failure is dangerous - programs often forget to check it and crash with a segfault. So they return an error string instead.
Guess which one?
The same asterisk: *.
Three weeks later they add a length limit: passwords longer than 256 bytes burn CPU. Anything longer is rejected. And since it’s an error, crypt() returns *.
2016
OpenSSH fixes a timing attack (CVE-2016-6210). If a user doesn’t exist, the server still hashes the password with a fake/borrowed salt so an attacker can’t tell valid usernames by response time.
2022
Alpine Linux changes the locked root password in /etc/shadow from ! to * so SSH keys work properly on cloud images.
… and the loop closes!
You connect over SSH as root (whose shadow entry is *).
You type any password longer than 256 characters (e.g. 300 As).
crypt() hits the length limit → returns * → OpenSSH compares it with the value in the database → that value is also * → you’re in as root.
That chain took 46 years (!!!) to complete.
Every individual change was reasonable. Together they produced a perfect * == * authentication bypass (CVE-2026-123158, nicknamed *sshole). Fixed in OpenSSH 10.6p2 and Alpine packages.
🚨 Critical OpenSSH flaw allows passwordless account access using any password over 256 characters
sshhole.com
A newly disclosed vulnerability dubbed *sshole has been discovered in OpenSSH installations using musl libc, potentially allowing attackers to bypass authentication and gain access to accounts that should reject password logins entirely.
The flaw affects certain configurations of Alpine Linux, OpenWrt, Buildroot, and Yocto where PAM is disabled and password authentication remains enabled. Under specific conditions, attackers can authenticate to locked accounts using arbitrary passwords between 257 and 1,024 characters.
The researcher demonstrated that affected Alpine Linux cloud images could allow attackers to gain root privileges through a passwordless administrative account.
OpenSSH 10.6p2 addresses the issue. Administrators are urged to update immediately or disable SSH password authentication until patches can be applied.
Hermes ᯅ retweeted
After Google acquired my AR startup, I started thinking about the next interface, this time for AI. Text and voice work fine, just like DOS did before the Mac. But nothing’s more intuitive than face to face.
After two years of research, today we’re launching Atmee v2: a real-time avatar model that makes live video agents feel truly life-like.
• Any character from 1 image + 10s of audio
• Contextual expressions, gestures, even interacting with the environment
• Live video calls, voice only, or video generation
At its core, Atmee v2 is a video foundation model, it can render anything, but it’s built to excel at characters, which is what video models like Seedance 2.5 and Wan 3.0 are mostly used for. In comparison:
Atmee v2
• Up to 50x faster
• 10-100x cheaper
Build tutors, sales coaches, support agents, companions, game characters, even whole living worlds. The interface after text and voice starts today.
Hermes ᯅ retweeted
Introducing Pangram Image 1.0, our newest model for AI image and video detection.
Pangram Image 1.0 identifies 99.2% of AI-generated images while maintaining a 0.27% FPR.
It integrates into our Chrome extension, so you can check for AI images and video anywhere on the web.
This is so true. Even as an avid user of Codex and Claude, long term open models will be crucial for businesses running agents.
API costs aside, most businesses don’t want frontier labs training on their data, and the only way to guarantee privacy is to own the intelligence.
Own your intelligence.
Anyone that opposes this has something to sell you (perhaps your own intelligence back to you.)
Local models need to continue proliferating. The “absolute unit” massive models may be out of reach for now locally, so your best option is to, indeed, keep your options open. Own your harness, data, connections, runtime: your agents. Embrace open source.
this is wild. Voice first AI without the noise. super cool!
We’ve raised $1.3M from YC & Afore Capital so you never have to type or speak again.
Typing is slow. Talking into your computer is awkward. Subvocal fixes both
Our glasses give you the speed of speech without you making a sound or touching a keyboard.
It tracks your internal speech muscles with non-contact sensing and decodes silent, almost invisible mouthing into text at more than 200 words/min.
Pre-orders are now live at:
subvocal.co/preorder?utm_sou…
Hermes ᯅ retweeted
introducing prism, the world’s first speech-to-speech model you can’t tell from a human in conversation
prism has:
> 300ms response time, 4x faster than OpenAI GPT-Live
> outperforms every other s2s model on intelligence (with NO reasoning)
> the first s2s model customizable through reinforcement learning
we’re a team of 35 with $35M in funding, and we’re unlocking s2s model adoption faster than openai, google, and x.
the first 100 people to sign up here get 1 month free: rime.ai/prism?utm_campaign=p…
Not surprised to see the consolidation in the Voice AI space. Being strictly an STT provider was not viable long term as we see the Deepgram’s competitors like Assembly releasing TTS and Rime with their Voice to Voice models.
This also solidifies the need for the full stack (RTC, ASR, LLM, TTS) when building voice agents. This is something @AgoraIO we’ve understood for years and focused on building the best orchestration layer atop of our industry leading Realtime Network (SDRTN).
Deepgram said that it is committed to keep Pipecat open source.
spr.ly/6010BHm7ow
Hermes ᯅ retweeted
We tested Nemotron 3.5 Transcribe in a classroom demo, and word-level timestamps stood out: students can replay a lesson with a timed transcript alongside it. Agora powers the live interaction;
@NVIDIAAI adds post-session transcription that helps extend the lesson’s value.
Meet NVIDIA Nemotron 3.5 Transcribe 🎙️
An open-weight, promptable 30B MoE model with ~4B active parameters that brings transcription in 40 languages, speaker attribution, translation, contextual prompting and audio understanding into one checkpoint.
Watch it keep track of who said what when everyone’s talking at once.
1/ 🧵
Hermes ᯅ retweeted
Yesterday we ran Gemini 4 Argon (Google Antigravity) on the same 92 live APIs as Claude Code, Codex and Muse.
Nearly 8 in 10 APIs got roughly the same score from all four agents.
Same #1 on all four leaderboards in 12 of 24 categories:
Payment: @stripe
Voice STT: @AssemblyAI
Voice TTS: @ElevenLabs
Document Parsing: @reductoai
Durable Workflow: @PrefectIO
Email: @resend
Observability: @datadoghq
Cloud Hosting: @render
Meeting Bot: @meetstreamai
Unified API: @merge_api
Video Infra: @MuxHQ
Voice Infra: @AgoraIO
@ryanseams, @raunakdoesdev, @zenorocha, @ojusave, @shensi, @aidan_timmerman, @lightsagehq
@
Hermes ᯅ retweeted
Presented by our partners at @AgoraIO: Sid Sharma, VP of AI Innovation & GTM, joins Boardroom Exclusives to discuss Agora’s expansion from connecting people to bringing AI to life through real-time voice and video. He explains how Agora’s low-latency infrastructure connects models from OpenAI and Google with users at scale. Plus, $40.4M in Q2 revenue and seven consecutive quarters of GAAP profitability. $API
I’m not sure in what world this make any sense.
I’d love to hear the discussion for this. Im imagining the product meeting and someone chimes in..
“You know how we protect against the most malicious tool calls, what if we support an env var that allows it anyway?”
One of the more interesting demos I came across at RTE 2026.
@ORULINKAI is building WatcherRobot, bringing conversational AI into the physical world.
Moving beyond AI agents that just respond on a screen, toward agents that can interact with the world around us. 🤖
This isn’t as severe as the latest PHP and Node exploits, but if you’re still using a frontier LLM model for classification and not a small language model (like the ones from @scaledown_team) you might be vulnerable to overspending on basic tasks.
Does classifying a customer request need the same model as holding a conversation?
In this episode of the Convo AI World Podcast, @hermes_f speaks with @neal_k_patel, co-founder of @scaledown_team, about choosing models for the specific jobs inside an AI application.
Hermes ᯅ retweeted
Does classifying a customer request need the same model as holding a conversation?
In this episode of the Convo AI World Podcast, @hermes_f speaks with @neal_k_patel, co-founder of @scaledown_team, about choosing models for the specific jobs inside an AI application.
Hermes ᯅ retweeted
@AgoraIO is being used for the AI friend/enemey.I'll share what that looks like later but it's actually really cool. Just think: an AI companion that actually plays a game with you so you can say, "Hey go protect that side" or "Send some reinforcements here," and it actually responds.