Offensive AI: Open Tools and Chain Vulnerabilities Disrupt Defense

Cascading intrusions linked to AI suppliers, open-source tools capable of automating attacks, and security teams under pressure: cybersecurity is undergoing a dramatic shift. Alex Stamos, Chief Security Officer at Cognition, describes a phase where AI enables real-time crossing of the "destruction chain" and challenges the economic balance of the sector. The benchmarks established since the early 2000s are being tested.

Benchmarks inherited from the Gates memo to the accessibility of attack methods

The current framework of software security was structured in the early 2000s around a call for trustworthy computing, a period when major companies like Microsoft and Oracle hired former hackers. Responsible disclosure practices, bug bounty programs, and consulting offerings became established. At the same time, the cybersecurity economy became more sophisticated: compensating specialists, including those encountered at DEF CON, to defend systems proved relevant. The sector debated the merits of obscurity versus open-source transparency, eventually setting standards for disclosure and notification practices. Alex Stamos argues that the economic landscape is changing with the accessibility of tools and models: a Mac Studio and an open-weight model would be enough to get started, without being socialized to professional standards.

AI-industrialized attacks, from prospecting to execution

According to Alex Stamos, adversaries' capabilities are escalating with AI, and open models now automate most of the destruction chain. This chain includes the steps that allow impacting a target, from information theft to system deactivation, as well as stealing money or Bitcoin. Where traditional methods relied on upstream research and human deployment, AI systems can now explore targets, identify bugs in real-time, write exploits in real-time, and penetrate companies. Alex Stamos also cites the use of AI by state actors, likening the current period to a new frontier, even going so far as to mention a transitional "dark age."

Software supply chain: from an AI supplier to 5,000 potential exposures

One link in the AI service provider chain, Fireworks, was targeted, and API keys were stolen. These credentials then allowed access to Cursor, where a malicious plugin was deployed. The extension affected 5,000 other companies, creating as many potential breaches. The Cursor incident is part of a series of attacks, with neoclouds targeted in succession and a hack of Binance presented as the largest known theft. Alex Stamos describes a climate of ongoing breaches within the tech industry.

Open tools, from ARTEX to GLM-5.3, and an already international reach

ARTEX, an open-source tool of Chinese origin, is accessible with an English translation on GitHub. Powered by a model like GLM-5.3, it could potentially facilitate the infiltration of a large number of organizations, according to Alex Stamos. Among the first victims mentioned are three South Korean banks, and according to this week's information, nearly 200 additional companies may have been affected. Alex Stamos emphasizes that it is remarkable that high-value entities can be compromised using freely downloadable tools, a situation that has not occurred in a long time. In this context, October, a month of security awareness and usual reminders of best practices, coincided with serious vulnerabilities in platforms like PHP and Node, along with multiple less-publicized intrusions. Security officials report operating at a heightened alert level since October and into November. The issue of AI-driven misinformation on social media is also among the concerns.