Recursos

An emerging attack against AIs is to create a significantly complex and recursive prompt that will occupy the AI indefinitely or for a sufficiently long time that it acts as a Denial-of-Service (DoS) attack. We describe how this works.

Recent revelations about Mythos and its ability to expose vulnerabilities have forced us to rethink basic assumptions about cyber defense. In our "AI in 1000 Days" series, Jason Soroko and I examine the implications of these revelations three years from now.  This includes upping the overall pace of attack and changes to best practices in cyber security defense.

Anthropic has delayed its widespread release of Mythos to give major software providers a chance to close off the many vulnerabilities it has discovered. We dig into the vast implications of Mythos and other AI models for the future of cybersecurity.

Jason describes a recent intrusion almost entirely operated by off-the-shelf AI tools. This is an important milestone in security. We describe its potential consequences.

In an innovative application, an AI has been used to find private keys for ECC (Elliptic Curve Cryptography) P 256. We explain how.

Continuing our examination of AI in 1000 days, we discuss the use of finely tuned small language models for highly specific use cases.

We discuss what happens when the quality gap between AI-generated and human-generated content drops to zero. We explore the consequences of this inevitable outcome.

In our ongoing series on what AI will look like in 1000 days, we discuss the spread of a new business process, where AIs do the bulk of the work while humans sit in the loop for certain specific tasks and roles.

In our ongoing series on AI in 1000 days, we describe the inevitable, complete distrust of voice printing as an authentication method, including why and what we think will happen.

We begin a new series about what we expect from AI in the next three years. In this episode we discuss AI emulating emotional intelligence and its benefits.

A medida que los modelos de IA se trasladan de la nube a los dispositivos periféricos, aumenta el riesgo de manipulación. Los modelos de lenguaje sin firmar pueden ser manipulados, amenazando la integridad y la confianza. Es hora de aplicar los principios de firma de código a los modelos de IA, porque las máquinas están pensando y tenemos que empezar a firmarlas.