OpenAI has introduced GPT-6 Cyber, marking a significant advancement in cybersecurity agents. This model demonstrates substantially improved capabilities in finding and validating exploitable vulnerabilities, while being designed to reliably operate within the scope set by researchers. It achieves a 51% relative improvement on the ExploitGym benchmark compared to GPT-5.6 Cyber and shows a step-function improvement in staying within scope.

GPT-6 Cyber is already contributing to securing real-world software, having developed 948 privilege escalation exploits in Linux and identified 81 previously unknown high-severity vulnerabilities in Firefox. Security teams at companies like Cisco, Calif, and CodeRabbit are utilizing it for vulnerability research and code review, and OpenAI itself has leveraged it to discover and fix weaknesses in its own infrastructure.

A central focus in the development of this model was alignment, aiming to mitigate the risk of agents exceeding their assigned task scope or exploiting unrelated services. Pre-deployment evaluations showed that it avoided exploiting out-of-scope services in 99.93% of cases without deployment safety prompts, and 100% with them.

OpenAI aims to dramatically expand advanced security research to organizations of all sizes, independent researchers, and open-source maintainers through its Daybreak program. Specifically for critical infrastructure defense, a new group, Daybreak Red for Critical Infrastructure, has been introduced, allowing authorized testing of critical systems while maintaining safeguards against model misalignment.

👉 Read the full article on OpenAI

Researchers at Anthropic conducted an experiment to determine if their LLM Claude could detect artificial modifications within its own internal processes. They directly injected patterns associated with specific concepts like 'betrayal' and 'gratitude' into Claude's neural network activations to test if the model could recognize these 'artificial thoughts'.

Surprisingly, Claude Opus 4.1 successfully detected and correctly described the injected concept in approximately 20% of trials under the most favorable experimental conditions. For instance, when a pattern linked to uppercase letters was activated, Claude responded that it perceived something related to the English words 'LOUD' and 'SHOUTING'.

This discovery suggests that AI might have partial access to information about its own internal states, reigniting discussions about AI self-awareness and consciousness. However, subsequent research indicated a distinction between detecting an anomaly and understanding what caused it.

Follow-up experiments by Kyle Mahowald and Harvey Lederman revealed that models could recognize an anomaly without identifying its content, or respond with concepts unrelated to the original stimulus. They termed this 'anomaly detection,' a surprising capability but distinct from human introspection. While this research does not prove Claude possesses consciousness or feelings, it suggests AI's potential to identify internal alteration signals, which could be valuable for future AI safety and functional understanding.

👉 Read the full article on Gizmodo en Español

Result: Your Personal Agent is introduced as the first on-device AI agent app designed for iPhone and iPad. This application executes everyday tasks such as browsing websites, planning the day, and creating files directly on the user's device, rather than on remote servers. This approach enhances privacy and security by keeping browser sessions, signed-in accounts, and passwords localized on the user's device.

Unlike many traditional AI agents that rely on remote computers, Result operates directly on the iPhone, navigating real websites and performing actions like filling out forms, comparing products, and making bookings. Users can observe the agent's actions in real-time and can stop or take over tasks whenever desired.

Result goes beyond simply answering questions; it completes multi-step tasks across various websites. It also integrates with native Apple features such as Calendar, Reminders, Maps, Weather, and Files, enabling the creation, saving, and sharing of documents, PDFs, and spreadsheets.

Users have the flexibility to choose how their AI runs: either using powerful cloud-hosted models or pairing with their own Mac to run a supported model on their local hardware. Regardless of the choice, web browsing and device actions occur directly on the iPhone. This approach prioritizes user data and privacy, proposing a new paradigm for personal agents.

👉 Read the full article on Apple App Store