A developer on Hacker News is asking a question that should worry anyone who stores passwords in a digital vault: if language models can attack systems with unlimited patience and iteration, how safe are password managers against an adversary running a swarm of AI agents?
The attack model is real
The poster's concern is specific and technically grounded. A criminal actor does not need a frontier model to attack a password manager. A mid-sized model looped until it finds a vulnerability, or a small swarm of agents exploring different attack vectors in parallel, can probe for weaknesses that human researchers would take months to discover. The advantage of AI attackers is not intelligence. It is patience and scale. They can try thousands of variations on a vulnerability class without fatigue, without getting bored, and without stopping.
The poster references exo, a distributed computing framework, running across two Mac Studios as a plausible attack platform. That is not a hypothetical threat actor with nation-state resources. That is a developer with a few thousand dollars in hardware and a language model that can generate and test attack payloads. The barrier to entry for AI-assisted attacks is dropping faster than the barrier to defense.
Where password managers are vulnerable
Password managers encrypt stored credentials, and that encryption is generally strong. The vulnerability the poster identifies is not in the storage layer. It is at the client and transport level. The client application that decrypts and displays passwords runs on your device. The transport layer that syncs passwords between devices crosses the network. Both are attack surfaces that AI agents can probe.
Client-side attacks target the application that handles decrypted passwords. A vulnerability in the password manager's code, a memory disclosure, a rendering bug that leaks data to the screen, or an injection vulnerability that lets an attacker execute code in the context of the password manager gives access to the decrypted vault. Language models can generate payloads targeting known vulnerability classes, test them against the application, and refine their approach based on the results.
Transport-level attacks target the communication between the client and the server. A certificate validation weakness, an authentication bypass, a session management flaw, or a side channel in the sync protocol could allow an attacker to intercept or modify password data in transit. AI agents can probe these paths systematically, testing edge cases that human testers might miss.
What the security community is actually doing
The password manager industry has invested heavily in security audits, bug bounties, and formal verification of critical components. Most major password managers undergo regular third-party audits. Many run bug bounty programs that pay researchers for discovered vulnerabilities. Some have pursued formal verification of their encryption implementations.
But these defenses are designed for human attackers. The assumption underlying most security testing is that an attacker has limited time, limited budget, and limited patience. AI attackers violate all three assumptions. A model that can generate and test attack payloads indefinitely changes the economics of vulnerability discovery. What was once a rare finding requiring expertise and persistence becomes a routine output of an automated process.
The honest answer to the poster's question is that password managers are probably safe today against the specific attack described, because the attacker still needs to find a real vulnerability in a specific application, and the major password managers have invested in making that difficult. But the trajectory is concerning. As models improve and attack tooling matures, the window between vulnerability discovery and exploitation shrinks. The defenders need to be right every time. The attackers need to be right once.
What users can actually do
The practical mitigations are limited but real. Use a password manager that has a strong security track record and undergoes regular audits. Enable multi-factor authentication on your vault, because it protects against client-side attacks that steal your master password. Keep your password manager updated, because security patches close vulnerabilities before they can be exploited at scale. And be aware that the threat landscape is changing. The tools that protect you today are designed for an era of human-speed attacks. The era of machine-speed attacks is arriving, and the defenses need to keep pace.
The poster's question is not paranoia. It is a reasonable assessment of a threat that the security industry has not fully adapted to. The password managers that survive will be the ones that assume their attackers have unlimited patience and test their own systems accordingly.