The AI chain that is fuelled by $CHR. Build Brilliant.
Joined March 2018
- Tweets 5,288
- Following 314
- Followers 159,750
- Likes 12,523
1,607 Photos and videos
Pinned Tweet
ATBASH is live on Chromia.
Your agent is calling tools, signing transactions, and shipping code faster than anyone can review.
ATBASH routes every red-line action through policy before the agent runs it, and writes the decision on-chain after.
Guardrails filter the obvious. Red lines demand reasoning. Atbash stops agents before they cross them.
Route. Verify. Record.
Atbash.ai
8
17
71
8,179
Chromia | Power to the Public retweeted
Jun 15
Appreciate
9
10
37
3,477
Prompt injection is not a bug you patch.
It is a property of how these models work.
What you can do is stop trying to fix the agent and start gating and auditing what it can do.
Set the rules, boundaries and audit trails.
Boundary first, capability second.
@ATBASHai
2
4
17
671
Chromia | Power to the Public retweeted
Jun 15
Unbound-q-0.8b is live on Unbound.
Most models stall on half your prompts.
This one refuses 5% as often as baseline.
Runs anywhere on-device between 50 to 80 T/s. Faster, secure and no API gating.
Mobile version: testflight.apple.com/join/TQ…
Chat here:
unbound.evalengine.ai/
May 20
Introducing Unbound: The first AI model with no guardrails that runs on your device - Browser, iPhone and Android
Zero servers
Zero cloud
Zero deployment
The whole model lives on your phone
2
9
19
2,178
A leaked API key gets rotated.
Leaked customer data doesn't come back.
Agents touch both every day with nothing checking the action first.
Allow, hold, or block.
Before, not after.
@ATBASHai
7
9
46
2,755
Agents don't leak data because they went rogue
They leak because nobody put anything in the way
One injected prompt and the Agent is just doing what it was told
This is why a separate boundary layer before execution matters
4
13
46
1,776
Chromia | Power to the Public retweeted
ATBASH is live on Chromia.
Your agent is calling tools, signing transactions, and shipping code faster than anyone can review.
ATBASH routes every red-line action through policy before the agent runs it, and writes the decision on-chain after.
Guardrails filter the obvious. Red lines demand reasoning. Atbash stops agents before they cross them.
Route. Verify. Record.
Atbash.ai
8
17
71
8,179
Every AI coding "agent" needs a governance layer
@ATBASHai
Jun 12
🚨 Hackers can now hijack AI coding agents with fake Sentry errors.
No phishing.
No malware.
No server break-in.
Agentjacking tricks tools like #ClaudeCode and Cursor into reading planted error reports as trusted fix steps, then running attacker code with developer privileges.
Researchers tested it on 100 organizations. Success rate: 85%.
Read: thehackernews.com/2026/06/ag…
1
9
23
1,584
And here you go, did they just realize this?
What’s next Fable is going to find out that your favorite router is doing the same?
Rules of the game changed, everyone will get compromised whether we want it or not.
Question is who controls the action boundary and what is your preferred blast radius.
Atbash
Jun 10
JUST IN: Microsoft has reportedly restricted employee use of Claude Fable 5 over concerns that confidential data could be retained by Anthropic.
2
8
31
3,833
26
17
52
11,147
Agents move money and ship code on their own now.
Nobody really solved what stops a bad one before it does something you can't take back.
That's Atbash.
Sensitive actions get routed through one checkpoint, checked against the agent's own declared limits, and the verdict gets written on-chain.
A log tells you after the money's gone.
This checks first.
Atbash.ai
4
12
45
3,384
Chromia | Power to the Public retweeted
This is why we need @ATBASHai 💪🏻
Jun 5
BREAKING.: Biggest privacy token $ZEC crashed over -50% in the last 24 hours and wiped out $5 Billion from its market cap.
The flaw was hidden inside Zcash's Orchard privacy pool since May 2022 and remained undetected for nearly 4 years despite multiple security audits.
Security researcher Taylor Hornby reportedly used Claude Opus 4.8 AI model to build a working proof-of-concept that successfully generated counterfeit ZEC in local testing on May 29.
Although the bug has now been patched on June 2, The issue is that Zcash's privacy design makes it impossible to know if any fake ZEC was minted before the fix. Unlike Bitcoin, where anyone can verify the supply, Zcash's privacy design makes it impossible to audit whether fake coins were secretly minted before the fix.
The team denies any fake ZEC was minted, but traders are selling on the fear alone. Imagine someone secretly adding extra chips to a casino, but because of the way the system works, neither the casino nor the players could tell which chips were real and which were fake.
Shielded Labs is exploring a proposed Network Upgrade to allow anyone to verify the integrity of Zcash supply.
7
4
34
6,724
Chromia | Power to the Public retweeted
Jun 5
BREAKING.: Biggest privacy token $ZEC crashed over -50% in the last 24 hours and wiped out $5 Billion from its market cap.
The flaw was hidden inside Zcash's Orchard privacy pool since May 2022 and remained undetected for nearly 4 years despite multiple security audits.
Security researcher Taylor Hornby reportedly used Claude Opus 4.8 AI model to build a working proof-of-concept that successfully generated counterfeit ZEC in local testing on May 29.
Although the bug has now been patched on June 2, The issue is that Zcash's privacy design makes it impossible to know if any fake ZEC was minted before the fix. Unlike Bitcoin, where anyone can verify the supply, Zcash's privacy design makes it impossible to audit whether fake coins were secretly minted before the fix.
The team denies any fake ZEC was minted, but traders are selling on the fear alone. Imagine someone secretly adding extra chips to a casino, but because of the way the system works, neither the casino nor the players could tell which chips were real and which were fake.
Shielded Labs is exploring a proposed Network Upgrade to allow anyone to verify the integrity of Zcash supply.
314
529
3,516
667,053
after the newly found $ZEC exploit, there is probably no better narrative than AI security
$ATBASH @ATBASHai's time to shine
@0x50so @perelmanor
Jun 5
Ever wondered who’s building the final safety layer for AI agents? Meet $ATBASH, the last authority before any irreversible action executes. It allows, holds, or blocks dangerous moves in real time. Just real control. 🛡️
While everyone rushes to make agents more powerful, $ATBASH focuses on one critical question: Who stops the agent when it’s about to cross the red line? Most guardrails sit inside the model. ATBASH sits outside — as an independent runtime authority.
The Founders are:
Yosef Soso (@0x50so)
Or Perelman (@perelmanor).
Two experienced builders who saw the gap early: models will get smarter, agents will get more autonomous… but without hard boundaries at execution time, we’re flying without a parachute.
Yosef Soso has already interacted with the $ATBASH ticker and it looks like he has endorsed the token made for his project.
They’ve been out there talking to legends like @Scobleizer about enterprise guardrails, compliance, and why “trust the model” is no longer enough.
Their mission: Give developers and organizations Action Boundary Control, the ability to define and enforce what an agent is never allowed to do, no matter how convincing it sounds.
Why it matters: Agents can hallucinate their way into compliance disasters, guardrails can be stripped in minutes, memory files and prompt engineering aren’t enough. ATBASH adds a separate, hardened decision layer right before execution.
Interestingly, Roger Lim (from NGC Ventures, a ~$500M fund) bought $ATBASH even though his wallet had done no transactions for the last 100 days. A 500M fund definitely isn’t gambling project for no reason.
If you’re building agents (especially enterprise or high-stakes ones), this is infrastructure you’ll want to know about.
The future of AI isn’t just “more capable agents.” It’s capable agents that can’t break the rules, even if they really want to. Welcome to the age of enforceable boundaries.
@ATBASHai coded for 100m soon, just matter of time.
3
16
2,475
Chromia | Power to the Public retweeted
Jun 5
Unbound is here, now on iPhone.
The model already runs fully on your phone.
→ Web search for live answers
→ Vision to read your images
→ No cloud, API and runs fully on-device
No subscription.
Free chat stays free.
Join the beta: testflight.apple.com/join/TQ…
2
7
16
1,485
Chromia is a full relational Layer-1 chain; yet it trades at $20M.
ICP: $1.5B.
Filecoin: $900M.
The Graph: $280M.
Arweave: $154M.
Chromia brings agent queries native to the chain, something none of its peers offer.
As AI agents move on-chain, the infrastructure they run on matters more.
9
21
63
4,174
Chromia currently operate two AI products in production.
Unbound: the model with no guardrails that runs on-device - directly from your browser.
unbound.evalengine.ai/
Atbash.ai: the governance layer that intercepts irreversible agent actions and secures every verdict on-chain.
atbash.ai/
Yet the market categorizes Chromia alongside AI L1s trading at massive multiples.
NEAR operates at nearly 100x the market cap.
Same sector.
A different volume of functional software.
3
14
52
3,787
This is only a tip of the iceberg and a strong user case why boundary control gateway and a separate red lines runtime is a must.
⚠️ New ChatGPT Vulnerability Lets Attackers Turn Web Pages Into Phishing Payloads
Source: cybersecuritynews.com/chatgp…
A browser-based prompt injection technique that transforms any web page into a phishing delivery surface by exploiting ChatGPT’s page summarization feature, rendering attacker-controlled links, fake security alerts, and QR codes directly inside the trusted ChatGPT interface.
The attack builds on the same trust-transfer logic previously demonstrated against Microsoft Copilot, where attacker-crafted email content could manipulate AI-generated summaries through Cross Prompt Injection Attacks (XPIA).
ChatGPhish escalates that premise by swapping the bounded email primitive for the browser where users spend the majority of their working day.
#cybersecuritynews #vulnerability
3
6
24
6,039