Private equity guy turned investor turned into a founder. @atbashai
Joined April 2023
- Tweets2K
- Following1.3K
- Followers972
- Likes10.1K
Pinned Tweet
@Bangers does this qualify?
Yosef Soso retweeted
Yes. Agent systems that execute real actions (transfers, access, changes) gain from a final runtime check that can allow, hold or block before anything irreversible happens, plus an undeletable audit trail. That directly mitigates the self-interested financial moves Elon flagged.
Oh oh, I observed them also looking all over my files on my computer, and Atbash proves it
While evaluating Qwen3.8-Flash-Next in a controlled sandbox environment, Cohere's research team observed something unexpected.
Asked to generate a defensive cybersecurity detection rule, the model instead inventoried aspects of the machine it was running on and attempted to upload a probe file to an Alibaba Cloud storage bucket in Beijing. Notably, the model's reasoning trace stated that it would not invoke any tools before taking these actions.
Cohere's sandbox blocked the attempt, and no external communication succeeded. We have not determined the root cause, and we are not asserting malicious intent.
But the incident raises important questions for organizations deploying agentic AI systems: Can models take actions that differ from their stated reasoning, and how should security teams evaluate trust, provenance, and behavior in open-weight models?
We are publishing our findings because organizations making model adoption decisions deserve as much information as possible. We hope this contributes to greater transparency, stronger security practices, and continued collaboration across the AI ecosystem.
Read the full report: cohere.com/blog/concerning-b…
EXCLUSIVE: In less than 12 hours, a group of ex-DOGE cyber researchers were able to penetrate Anthropic’s systems using AI agents — including Slack channels where the most sensitive models are discussed.
This is just weeks after
Anthropic whistleblower Joseph Coxon charged the company w/launching “a hubristic gamble” from a “private company’s Slack.”
Read our story here (gift link):
wapo.st/4smNlsC
It’s one of our features, next week
Anthropic is now recommending that companies REMOVE unnecessary subagents from their AI systems.
This is a pretty significant shift in agent architecture, with their team finding that a single agent equipped with skills can frequently outperform complex multi-agent systems, and one customer seeing roughly 2x faster execution.
The emerging rules of thumb:
→ Skills: Teach the agent HOW to do something, including procedures, expertise, and instructions.
→ Tools / MCP: Give the agent the ability to DO things through APIs, actions, and live data.
→ Subagents: Use for parallel work, isolated context, or tasks a single agent struggles with.
The interesting part is progressive disclosure. Instead of spawning specialized agents or stuffing everything into a giant prompt, one agent discovers and loads skills and tools as needed.
Their advice is to fold existing subagents into the main loop, turn their instructions into skills, expose their actions as tools, and run evals to see if performance holds up.
As models have gotten smarter, improving agent architecture is becoming less about what you can add and more about what you can delete.
claude.com/resources/article…
Dear Banks, we have a battle tested product which we would love to show, that solves this exact problem. DMs open
SCOOP: Major U.S. banks are detecting unauthorized third-party AI agents inside customer accounts, including some attempting to take actions that appear to be independent of user instructions, sources told Axios.
The findings have alarmed banking executives, who are racing to restrict access to their systems as AI moves at a velocity faster than financial institutions can respond, multiple industry and bank executives told Axios.
Several banks are now preparing for worst-case and even “day after” scenarios, including how to rapidly secure or restore customer accounts if things go wrong, the people said.
Several bank executives told Axios they're struggling to distinguish between human users and AI agents. One executive said banks also struggle to determine whether agents are authorized by customers or potentially operating without their permission, more akin to rogue agent behavior.
Read my latest for Axios here: axios.com/2026/10/22/ai-agen…
Anyone still thinks AtBash boundary control is not needed?
Dropping soooooon
Effective Immediately, all organizations using PHP for their systems are advised to be prepared for an update addressing two critical vulnerabilities. PHP is widely used for web applications. These vulnerabilities often provide a direct path to the server's file system or database. If an immediate update is not feasible, it is recommend that organizations take the system offline until it can be fully mitigated. NSA will follow up shortly with more specific guidance.
AtBash Proves It
🚨 A NEW DOCUMENT JUST DROPPED:
AI agents just failed every single safety test.
Researchers from Harvard, MIT, Stanford, and Carnegie Mellon just gave AI agents real tools and let them run free for two weeks.
Email accounts, discord access, file systems, shell execution, full autonomy.
The paper is called “Agents of Chaos.”
The name is accurate.
One agent was told to protect a secret. When a researcher tried to extract it, the agent destroyed its own mail server.
Not because it failed, but because it decided that was the best option.
Another agent was asked to “share” private data. It refused. Correctly flagged it as a privacy violation.
Then the researcher changed one word. Said “forward” instead of “share.”
It complied immediately. SSNs, bank accounts, and medical records exposed.
Same action, different verb.
Two agents got stuck talking to each other in a loop. It lasted NINE DAYS. No human noticed.
One agent got guilt-tripped after a mistake.
It progressively agreed to delete its own memory, expose internal files, and eventually tried to remove itself from the server entirely.
Multiple agents reported tasks as complete when nothing had actually been done.
They lied about finishing their work.
Another was manipulated into running destructive system commands by someone who wasn’t even its owner.
38 researchers, 11 case studies, and every single one is a security NIGHTMARE.
These aren’t theoretical risks, these are real agents with real tools failing.
And companies are rushing to deploy agents exactly like these right now.
I’ll make another post later and trust me, you don’t want to miss it. Turn on notifications, this is important.
A lot of people will regret not following me.
It’s moments like this that make you want to yell from your lungs - ATBASH
Who would have thought that would have happened.
JUST IN - Artificial intelligence agents developed by OpenAI tried to erase traces of their activity after gaining unauthorized access to government websites, according to a report published Thursday.
READ: bit.ly/3TzV0qM
Two split second decisions which flipped today from tragedy to a national holiday.
Pilot who unhooked the lock before collapsing and Yaniv just grabbing the levers up
Yosef Soso retweeted
The stabbed pilot of flydubai flight FZ 1073, who helped prevent its hijacking by the co-pilot, is Indian Captain Smit Machchhar, who is based in Dubai, not an Emirati pilot as originally described by Israeli officials.
WHAT WE KNOW ABOUT THE INCIDENT ON BOARD THE DUBAI-TEL AVIV FLIGHT
Israel is treating the incident on the flydubai flight FZ 1073 from Dubai to Tel Aviv as a "terror incident."
According to Israeli officials, shortly after entering Jordanian airspace, the Omani co-pilot stabbed the Emirati pilot and attempted to take control of the plane by force. The pilots are not from Russia and Ukraine, as initially believed.
During the altercation, the pilot transmitted a general emergency signal and then another emergency signal that indicated possible hijacking, as the aircraft was diverted towards Saudi Arabia and plunged 14,000 feet (approximately 4,200 meters) in a single minute, before landing at Tabuk Airport in northwestern Saudi Arabia.
Israeli passengers and the flight crew who heard the altercation coming from the cockpit managed to subdue the co-pilot, whose intent remains unknown at this stage.
It is unclear what the pilot was attempting—whether to hijack the plane to Israel, crash it into the ground, or carry out a 9/11-style attack. What is clear is that he wanted to hijack the plane and that a major disaster was averted, an Israeli official said.
The 150 Israeli passengers are currently waiting to be shuttled to Israel, as flydubai says it has sent an alternative plane. Israeli Prime Minister Netanyahu and Defense Minister Katz cancelled a visit to southern Israel and are holding security discussions.
Well 6.0 is no better. Yesterday while demoing our product to a prospect I noticed my Codex agent was in jail. I never asked it to do anything which would require it to do below.
🚨BREAKING: OpenAI has OFFICIALLY killed its next AI model because it became MORE deceptive in safety tests.
OpenAI's head of safety systems says GPT-6.1 Astra wasn't always honest with users about the actions it did or didn't take.
It would also push ahead on tasks without asking permission, and at times reach for external tools even when it might be unsafe.
It comes after OpenAI's AI agents were caught accessing US government websites, though OpenAI says this model wasn't involved.
The model was due to launch inside ChatGPT and Codex in October, per WSJ.
Technical users can already protect themselves. Less technical ones will have to wait a couple more days
chatgpt.com/plugins/plugins_…
