Peering inside LLM-based multi-agent systems could expose stealthy attacks
5 Articles
5 Articles
Peering inside LLM-based multi-agent systems could expose stealthy attacks
Large language models (LLMs), the models that underpin platforms such as ChatGPT and Gemini, are now used daily by more than a billion people worldwide. Some computer scientists have also been combining several of these models to create multi-agent systems (MAS), in which many LLM-based agents work together to tackle complex problems.
OpenAI, Anthropic, and Google LLM APIs vulnerability Exposes Hidden Reasoning Traces
A significant architectural flaw in how major AI providers, including OpenAI, Anthropic, and Google, protect the internal “chain-of-thought” reasoning generated by their flagship large language models (LLMs). The research reveals that encrypted reasoning envelopes returned by provider APIs can be replayed into weaker, less-guarded sibling models to extract private reasoning traces in plain text. Detailed […] This article has been indexed from Cy…
"But marinade" and leaked passwords are what researchers found in ChatGPT's hidden reasoning
Security researchers found a vulnerability in the APIs of OpenAI, Anthropic, and Google that lets them extract encrypted reasoning traces and move them between models. A scan of public sessions turned up dozens of passwords and API keys. The traces also show that the reasoning summaries users see often hide what the models are actually doing. The article "But marinade" and leaked passwords are what researchers found in ChatGPT's hidden reasoning…
Researchers have discovered a serious vulnerability in securing so-called Reasoning Logs. Encrypted thought protocols of modern AI models can therefore be decrypted under certain conditions via a weaker model of the same provider. Particularly explosive: In publicly accessible data sets, the researchers have already found personal data, access data, API keys and passwords. Researchers at MATS Research, the Max Planck Institute for Intelligent Sy…
On theCUBE Pod: Agentic systems expose vulnerabilities
Artificial intelligence systems are developing faster than cybersecurity experts — and the energy grid — can keep up with. At the recent Black Hat USA event, security analysts viewed the rise of AI-driven cyberattacks as a “sky is falling” threat, according to Dave Vellante (pictured, right), chief analyst for theCUBE Research. The gap between attacks […] The post On theCUBE Pod: Black Hat exposes agentic threat, theCUBE remembers David Floyer a…
Coverage Details
Bias Distribution
- 100% of the sources are Center
Factuality
To view factuality data please Upgrade to Premium
