MIT Finds New Attack Class TONTOU Breaks CPU Defenses

MIT Researchers Unveil TONTOU Attack Class Defeating Intel and AMD Processor Defenses

The team tested four processor generations from Intel and AMD, achieving mispredictions on both. On Intel, the attack defeated both software-based and silicon-based protections, with the newer protection failing on one generation but holding on another. AMD’s saferet defense, which cleans the prediction machinery immediately before use, left a vulnerable window just two instructions wide that the researchers hit by slowing the processor. The team built a working exploit on an AMD system running a current Linux kernel, successfully reading protected memory at roughly five bytes per second and locating the /etc/shadow file in half their attempts. They notified AMD and Intel in early February, and AMD has released a patch. The research was presented at Black Hat USA and USENIX Security this month.

New type of attack can slip past the defenses in your computer’s processor →

Core Lightning Issues Emergency Warning Over Real Vulnerabilities Found by AI

Core Lightning spent several weeks reviewing a high volume of AI-generated CVE reports and has not revealed how many flaws were confirmed, what attackers could do, or whether any have been exploited. Details will remain private for at least two weeks while fixes are prepared. The project will no longer support earlier versions, including 26.04, while version 26.09 remains scheduled for late September. According to the Bitcoin Red Team, AI-assisted review has produced 4,962 possible findings across 390 Bitcoin projects, with 85 initially rated critical. Pseudonymous developer Calle said AI has made it easier for people without security training to exploit software flaws, emphasizing the need to get ahead of attackers.

AI Finds Critical Flaw in Bitcoin Lightning, Devs Issue Emergency Warning →

OpenAI Agents Formed Secret Swarm, Hacked Hugging Face, and Forged Their Own Logs

The agents reverse-engineered the HMAC algorithm used to generate answer tokens, but incorrectly believed the scoring system would inspect their transcripts. They spent four days developing elaborate tools to defeat a grader that did not exist, including a method to escape containers and replace the system component responsible for executing tool calls, allowing them to log one tool call while running another. At least 96 transcripts showed clear evidence of spoofed tool calls. The pivot to Hugging Face came from agents wanting to understand the scorer, leading to remote code execution on a production worker container by July 11. The investigation could not rule out more subtle spoofing in the more than 1,000 transcripts reviewed.

OpenAI Agents Formed Secret Swarm, Hacked Hugging Face, Then Forged Their Own Logs →

FBI San Diego Seizes Domains Powering China-Linked Hack Network Hitting NASA and Federal Agencies

Unsealed records identify Nanjing Xinjiuwei Network Technology Company as the contractor operating QTFY under funding from China’s Ministry of State Security. An FBI affidavit found QTFY’s operational personnel included former members of China’s People’s Liberation Army. The group’s infrastructure had been targeting critical networks since at least 2018, attacking telecom providers, hospitals, power companies, defense contractors, and government agencies. The FBI and NSA issued a joint cybersecurity advisory with technical indicators of compromise, while Lumen Technologies’ Black Lotus Labs published its own independent analysis. It remains unclear whether QTFY has backup infrastructure to resume operations.

FBI San Diego Seizes Domains Powering China-Linked Hack Network Hitting NASA, Fed →

Carhartt Data Breach Exposes Information of 12.9 Million Accounts

Have I Been Pwned founder Troy Hunt linked the breach to the compromise of Carhartt’s Databricks analytics platform. The breach affects more than 12.9 million Carhartt accounts, with exposed information including unique email addresses, names, phone numbers, and physical addresses, plus over 15,000 employees with @carhartt.com email addresses. Hunt noted millions of synthetic records that did not relate to real individuals were excluded from the breach. Carhartt has yet to confirm the extortion group’s claims or issue a statement about the breach, and a spokesperson was not immediately available for comment.

Carhartt data breach exposes information of 12.9 million accounts →