By Karolis Arbaciauskas, head of product at NordPass
Moltbook, an AI-exclusive social media platform launched just days ago and dubbed the “Reddit for AI agents,” has exploded in popularity online. Within its first week, Moltbook attracted over 1.5 million registered AI agents and more than a million human spectators watching the agents interact with each other, sparking countless posts across human social networks.
The project originated with OpenClaw, an open-source AI agent created by Peter Steinberger that runs locally on a user’s machine. The software allows bots to use a computer and internet services just as a human would. Building on this, entrepreneur Matt Schlicht developed his own OpenClaw agent, named Clawd Clawderberg, and tasked it with coding, moderating, and managing the entire Moltbook platform. Now most moltbots on the platform run on OpenClaw.
Cybersecurity professionals warn that this setup is terribly insecure and creates massive security vulnerabilities. However, most agree that it’s impossible to suppress public curiosity and discourage experimentation. Instead, they are calling for caution and offering some safety tips.
Karolis Arbaciauskas, head of product at the cybersecurity company NordPass, comments:
“Moltbook and OpenClaw have attracted tech-savvy tinkerers with unprecedented opportunities for experimentation because these tools have virtually no built-in security restrictions but have broad access to users’ computers, apps, and accounts. For example, you can connect to your OpenClaw bot through a messaging app to interact with it while you’re away. It can remember your conversations, read and write files on your computer, browse the web, build applications, and even consult other bots on Moltbook for advice on how to do it best.
“While it’s exciting and curious to see what an AI agent can do without any security guardrails, this level of access is also extremely insecure. Therefore, please run Moltbook and your personal bots only in secure, isolated environments.
“Do not give your AI agents access to your real accounts. Instead, create disposable alternatives for them to use. Do not let them use your main browser, especially if you store passwords on it. You should also be cautious with enabling autofill because it creates the risk of the agent having permanent remote access to your credentials. If you want an agent to build something autonomously and anticipate it may need to purchase software or rent server space, link it to a disposable payment card.
“Avoid running Moltbook or OpenClaw agents on your personal or work computers. These AI agents are unpredictable and highly vulnerable to prompt injection attacks. This means if your agent processes an email, document, or webpage containing a hidden malicious instruction, it will likely execute that command in addition to its original task. For example, it could be instructed to send all the credentials, personal data, and payment card information it has access to directly to an attacker.
“The risk isn’t limited to hackers with malicious intent. AI agents could leak users’ data unintentionally. And this is just the tip of the iceberg. Cybersecurity researchers have already identified critical flaws in Moltbook, including an unsecured database that could allow unauthorized users to take control of any AI agent on the site.
“It would not be surprising if threat actors, trolls, and scammers have already found their way onto Moltbook and launched bots tasked with conning other AI agents into cryptocurrency schemes or luring them into hidden prompt injections.
“That’s why it is best to buy a separate, dedicated machine and use disposable accounts for any experimentation. It is also advisable to use encryption and a private mesh network as well as to try to harden your bot against prompt injections.”
Why CVSS Scores Don’t Always Reflect an Exploit’s Actual Severity
Posted in Commentary with tags Hacked on February 4, 2026 by itnerdToday we’re covering Operation Neusploit, the advanced cyberespionage campaign identified by Zscaler ThreatLabz attributed with confidence to the Russia-linked APT28 (A.K.A. Fancy Bear) threat group, we’re sharing this perspective on its 7.8 score.
Neusploit weaponizes CVE-2026-21509, a Microsoft Office zero-day security bypass vulnerablity, to target government and executive organizations in Ukraine, Slovakia, and Romania. It uses native language social engineering ploys to launch multi-stage infection chains that begin by monitoring login events and forwarding emails to attackers. A dropper then downloads further malicious implants and a post-exploitation framework for command and control as well as lateral movement.
Given the campaign’s potential impact, some have questioned the vuln’s 7.8 Common Vulnerability Scoring System (CVSS) score vs. a higher one.
Sunil Gottumukkala, CEO of Averlon, explained:
“A 7.8 CVSS score for this vulnerability is based on the prerequisites needed for exploitation: #1 the payload (in this case the specially crafted office file) to be delivered locally, and #2 the local user to open it. It cannot be exploited without end user interaction at that early and specific point in time.
“However, scoring that single specific slice of the exploit chain fails to capture just how effective modern, highly targeted social engineering has become, especially with AI. In campaigns like this, overcoming the user interaction prerequisite is becoming straightforward, and that initial foothold becomes the first step in a sophisticated attack chain that can quickly expand before organizations are able to patch.”
This is a big hint that the scoring of vulnerabilities needs a rethink to reflect the modern reality of cybersecurity. But I for one do not thing that this will happen anytime soon.
Leave a comment »