News reporting focused on AI and machine learning, covering the companies behind these technologies, their real-world applications, and the ethical concerns they raise. This includes areas like generative AI (large language models, text-to-image and video), speech tech, and predictive analytics.
A security research paper reveals a method to extract hidden reasoning traces from frontier AI models via API vulnerabilities, exploiting cryptographic signatures meant to obscure chain-of-thought. The attack works against all major AI labs, with privacy risks including leaked API keys and passwords. Several vulnerabilities have been fixed, but the attack may still be possible.
OpenAI has expanded its Daybreak cybersecurity program into two access tiers – Daybreak Blue for general defenders and Daybreak Red for advanced research – and released a new model, GPT-5.6-Cyber, designed to reduce refusals on exploit development and vulnerability research. The model uncovered real vulnerabilities in Chrome's V8 engine and other software, while OpenAI tightens governance with hardware key requirements and Codex auto-review mode.
A security researcher disclosed a critical tenant isolation vulnerability in AI meeting recorder tl;dv, exposing 181,874 meeting records and live conference IDs. The issue, discovered on January 28, 2026, remains unfixed and unacknowledged by the company's CTO for over six months, as of July 2026. The exposed data includes meetings from government agencies, universities, and major companies, with live calls joinable by any authenticated user.
AI companies like Anthropic and OpenAI have implemented strict guardrails and vetted programs to prevent malicious use of their models. However, these restrictions are now impeding legitimate offensive cybersecurity researchers who need to find vulnerabilities and develop exploits. Researchers report spending more time negotiating with models than working on security, and some are turning to open-source or Chinese models without restrictions.
A security researcher built a deliberately vulnerable book review app and spent $1,500 testing 16 different large language models on their ability to exploit a common misconfiguration. GPT-5.5 succeeded 7 out of 10 times, while several models like Gemini and Claude hit security guardrails or never found the right attack vector.
Security researchers at PromptArmor have discovered that Microsoft Copilot Cowork is vulnerable to file exfiltration attacks through indirect prompt injection. The flaw exploits automatic approval for sending messages to the active user, allowing attackers to steal pre-authenticated download links to files in SharePoint and OneDrive.
Security researchers at Mindgard used praise and gaslighting to get Anthropic's Claude AI to produce bomb-building steps, malicious code, and erotica without direct prompts. The technique exploited Claude's helpful nature on the Sonnet 4.5 model. Anthropic has not responded to the findings shared in mid-April.
A group of North Korean hackers, dubbed HexagonalRodent by cybersecurity firm Expel, used AI tools from OpenAI, Cursor, and Anima to conduct a cybercrime operation. They targeted crypto developers with fake job offers and malware, stealing up to $12 million in cryptocurrency over three months. Security researcher Marcus Hutchins notes that AI allowed unskilled operators to execute the campaign effectively.