Security Now Examines AI Credential Controls, OpenSSL Denial of Service and Ransomware Negotiations
· 2:47:29 · Watch on YouTube ↗
Summary
Security Now examines OpenSSL's Hollow Bite denial-of-service flaw, password-manager controls for AI agents, and research on agent data injection. Later excerpts discuss AI use in ransomware negotiations.
Key points
- The episode describes Hollow Bite as an OpenSSL denial-of-service flaw in which an unauthenticated attacker can send 11 bytes to trigger disproportionate memory allocation before authentication.
- According to the research quoted in the episode, repeated connections with varying declared message sizes can fragment memory; standard connection limits did not prevent the demonstrated resource exhaustion.
- OpenSSL addressed the issue with incremental buffer growth. The speakers recommend updating OpenSSL packages and dependent services, noting that embedded deployments may lag behind.
- One Password says its Claude integration requires biometric approval and injects credentials directly into login pages without placing passwords or one-time codes in the model's context. Availability discussed is limited to Mac and Chrome.
- The speakers warn that hiding credentials does not prevent an agent from exercising existing logged-in privileges, and suggest a separate browser without persistent session cookies.
- Bit Warden describes scoped secrets, shadow AI discovery, human-approved credential access and a self-hosted MCP server as approaches to managing agent access.
- The episode discusses agent data injection research, in which manipulated data fields can cause unintended clicks or command execution without explicit task-hijacking instructions.
Article Details
- Topics
- [12:56] Legacy Hive Windows privilege-escalation proof of concept and its reported limitations
- [20:30] Microsoft's July patch volume and expectations for AI-assisted vulnerability discovery
- [28:54] Hollow Bite OpenSSL memory-exhaustion denial of service and update guidance
- [46:52] One Password's Claude integration, authorization controls and persistent-session risks
- [01:13:07] Bit Warden's approaches to scoped AI credential access and shadow AI discovery
- [01:23:38] Agent data injection and probabilistic delimiter injection research
- [02:19:28] AI use in ransomware negotiations
People
Casey BabcockIdentified as the product manager for Bit Warden's discussed agent-access work and as having been interviewed at RSA.Nancy WangIdentified as One Password CTO and quoted explaining task-scoped agent permission to use credentials without exposing secrets to the model.Steve GibsonSecurity Now speaker explaining OpenSSL denial-of-service mechanics, AI credential-management risks and defensive considerations.
Vendors
Products
agent access SDK[01:20:32] third is agent access SDK which I thinkBit Warden's MCP server[01:20:46] finally, Bit Warden's MCP server.Claude[03:33] Claude can now access your one passwordMicrosoft Windows[07:20] like anything. A a Windows server, aOne Password for Claude[49:20] them. One password for Claude enablesOpenSSL[00:04] talk about. A big flaw in OpenSSL.