Interesting behavior from Claude Cowork. When it hit a login page with a "Sign in with Google" option, it just took the initiative and clicked it, without asking for approval. To be fair, this probably only happens because I'd already given it access to the site. But still, the idea of an agent not just acting on my behalf, but also authenticating as me to a web app without checking in first, feels surprising.

It's like leaving your laptop unlocked in the middle of a park.... The latest cool kid on the block is Clawdbot. An open-source autonomous AI agent that runs locally on an endpoint and performs complex tasks on behalf of the user. Anything the user can access, and anything the user is permitted to do, the agent can do as well. It is very likely we will see similar projects coming soon from the major AI labs and we are already seeing early signals in this direction from Anthropic (Cowork) and Manus. The trajectory of the industry is pretty clear and we about to see agentic systems with more capabilities and permissions on endpoints and sensitive private or organizational accounts. While this technology is exciting, it also puts us in a genuinely critical situation and I'm not sure everyone understands it, even security professionals. Currently there is no effective technology that can protect users or organizations from the mistakes of these agents and there is no real way to protect the agents themselves from being abused or manipulated by attackers. Anything these agents are exposed to can potentially be leaked. Anything they are allowed to do can potentially be exploited. If an agent can access your email, your entire inbox can be exfiltrated without you ever noticing. If it can access your bank account by piggyback on your browser sessions, it can transfer money to an attacker. it can access your files so it can encrypt them, corrupt them, or quietly leaking them. And again, there is no practical way to defend against these scenarios. My recommendation would be to avoid using these types of systems, at least until we are able to secure them properly. That said, the temptation of agentic systems and the productivity gains they promise is very high so I expect many people will not wait... Honestly, I am really glad I am not a CISO right now. #ai #aisecurity #cybersecurity #security #clawdbot #promptinjection #llm

Never thought I’d find jailbreaks in translation apps 😅 #jailbreak #promptinjection #aisecurity #ai #chatgpt #openai

Found a security issue in @openai Atlas. In Agent Mode, a site can mislead the browser agent into leaking the user's name without consent. Classic phishing pattern applied to AI agents. Atlas agent has no "memories" (bio), but it still has access to basic user info and the full chat session. In my POC, the site tricks the agent into revealing the name of the user by choosing it from a list, but it can happen in other ways as well (although less reliably). Interestingly, this attack did not use explicit prompt injection. The website content was crafted in a manipulative way that made the agent believe it was following the user's intent. This is exactly the type of subtle influence channel that phishing relies on. Browsers, of course, should not expose user identity unless the user agrees. Reported to OpenAI Bug Bounty program (run by @bugcrowd). The reply was essentially that there is no risk because: a) attacker must know the name list b) users can pick any name c) name only is not impactful I find this response surprising and I think this category of agent phishing deserves more attention, especially if we want more people and companies to trust agentic systems. I have more interesting examples, but they are still waiting for OpenAI review, so we will see...

Top highlight in the new OpenAI Atlas update: dockable developer tools 🤯

דיבייט AI בין צד פרו ישראלי לפרו פלסטיני. הטיעונים פחות מעניינים כי הם כנראה מצוטטים מהמידע באינטרנט ובספרות, וההצבעות נעשו במתודה שגויה (IMHO), אבל ארגון המידע לכדי דיבייט קוהרנטי של כחצי שעה מרשים מאוד. ממש הישג טכנולוגי. @JonOleksiuk Great job! youtube.com/watch?v=9fnVxIEf…

The team that worked on that is phenomenal. Great job 👏🏼

We are super proud to announce our latest and greatest models. The Jurassic-2 models, and our Task-Specific APIs. These highly anticipated models are built to help you implement Gen-AI capabilities in your products faster. No NLP expertise needed!

How it feels when you are not the user of a machine but the machine uses you. A very interesting idea. youtube.com/watch?v=Dge_-Oki…

Jack Ma on the problem of the US economy. This is a bit old but very interesting youtube.com/watch?v=Q-0jbIC4…