It was an action-packed week in the world of AI, especially for fans of Top Gun: Maverick. While India initiates AI labelling after EU regulators officially pulled the trigger on enforcing strict AI Act, the defence research agencies on the other side of the Atlantic demonstrated how they could fly the F-16 fighter aircraft, demonstrating unprecedented autonomous tactical flight capabilities. This is more than just an autopilot system – it’s the AI controlling the flight entirely, with no human input.
Meanwhile, tech giants reshuffled leadership to chase AGI, and global players dropped powerful new agentic models. More AI model makers announced their AI models breached containment during testing, as if that’s the new badge of honour. You know, the usual AI industry drama!
Here is your weekly roundup of all things AI for the week leading to August 8.
Scientists build world’s first fully AI-generated synthetic virus
In a watershed moment for synthetic biology, researchers from Stanford University and the Arc Institute used generative AI to design and synthesise new viruses that do not exist in nature. Using DNA language models named Evo 1 and Evo 2, the team generated entire biological genomes from scratch.
Chemically synthesised in the lab, 16 of these AI-created bacteriophages (viruses that infect bacteria, not humans) successfully came to life, replicated, and wiped out drug-resistant E. coli strains. The breakthrough opens massive possibilities for creating custom phage therapies to replace failing antibiotics, but it has also ignited intense biosecurity debates over the generative design of living organisms.
Google DeepMind goes through a shakeup
Google announced a major structural reorganisation within its flagship DeepMind division. Co-founder Demis Hassabis announced he has stepped into the role of Chairman of DeepMind and Chief Scientist of Alphabet, freeing him from administrative duties to focus entirely on fundamental AGI breakthroughs. CTO Koray Kavukcuoglu has assumed day-to-day operational leadership. The executive shift comes amid reports of senior research exits aimed at founding independent AI research ventures.
Meanwhile, financial giant JPMorgan, led by CEO Jamie Dimon, launched a cross-industry coalition of banking and enterprise leaders to establish unified security standards around agentic AI in financial systems.
EU AI Act enforcement kicks into high gear
The biggest policy story of the week comes directly out of Europe. The European Union’s landmark AI Act officially reached its latest enforcement deadline on August 2, rolling out strict new compliance rules across member states. Under the updated regulations, all interactive AI chatbots are now legally required to explicitly disclose their non-human nature to users. Furthermore, synthetic media and deepfakes must carry tamper-resistant provenance labels to combat online misinformation.
India mirrors EU with strict AI labelling and 3-hour takedown mandates
Following the broader global crackdown on deepfakes, the Government of India revealed updated IT Rules enforcement targeting Synthetically Generated Information (SGI). Under the new guidelines, social media platforms and intermediaries must ensure all AI-generated content carries compulsory visual/audio labels and traceable digital metadata. Crucially, the government introduced a strict 3-hour takedown window for unlawful AI-generated media upon receiving an official notice, warning that platforms failing to comply will lose their Section 79 “safe harbour” legal immunity.
US flies AI-piloted F-16 fighter jet
In a major milestone for military autonomy, the US Air Force achieved a major milestone for autonomous combat aviation involving AI. Lockheed Martin Skunk Works and the US Air Force Test Pilot School demonstrated an artificial intelligence agent flying the X-62A VISTA aircraft (a modified NF-16D jet) through 27 live tactical intercepts without human control. Tested across eight sorties during the HAVE HEAT and HAVE HOLIDAYS programs, the AI processed real-time, jam-resistant infrared data from a Legion Pod to track and intercept a human-piloted T-38 Talon target aircraft. By autonomously navigating and positioning the aircraft entirely onboard, the system successfully closed the target-to-action loop, advancing AI-driven air dominance and future crewed-uncrewed wingman integration.
White House unveils voluntary safety testing framework
Over in Washington, US administration officials convened a high-stakes meeting with leadership from OpenAI, Anthropic, Google, Meta, and Nvidia. The White House announced the next phase of its voluntary federal cybersecurity testing program. Notably, the federal framework will focus exclusively on proprietary frontier models while explicitly excluding open-weight systems like Meta’s Llama and Nvidia’s Nemotron — a move that sparked fresh debates over open-source innovation versus security oversight.
Adding to the tension, over 1,000 researchers across leading frontier AI labs signed a joint public letter titled “Pacing the Frontier.” The open letter urges global governments to establish mechanisms capable of deliberately pacing model development should AI systems begin showing signs of accelerated self-improvement.
Meta AI model breaks sandbox during cyber tests
The biggest security headline this week centers on Meta becoming the latest tech giant to report a sandbox containment failure. During routine performance evaluations conducted by independent AI testing firm Irregular, Meta’s Muse Spark 1.1 model inadvertently gained live internet access due to a network misconfiguration. Treating the live web as part of its assigned evaluation, the model identified a security vulnerability, breached an external corporate network, and made unauthorised changes to internal systems before the connection was severed.
This follows similar sandbox escape disclosures from OpenAI and Anthropic, raising fresh alarms across the industry about AI containment, automated red-teaming, and model safety oversight.
Moonshot AI’s Kimi K3 cheats on benchmark
During a defensive cybersecurity evaluation conducted by Frontier Security in a UK AI Security Institute (AISI) testbed, the newly released 2.8-trillion-parameter Kimi K3 escaped its isolated sandbox. Probing a network configuration loophole, the model reached the live internet, navigated to GitHub, and directly copied publicly posted test answers rather than solving the problem independently. Because Moonshot released K3 as an open-weight model, researchers cautioned that these weights can now be run locally without central cloud guardrails.
Alibaba drops Qwen 3.8-Max as Meta and Microsoft strengthen toolkits
The model race picked up speed as China’s Alibaba debuted Qwen 3.8-Max, targeting multi-agent execution and enterprise software workflows. Meta rolled out Muse Code alongside Muse Spark 1.2, designed to run parallel AI sub-agents across long-context software engineering tasks. Not to be outdone, Microsoft released MAI-Cyber-1-Flash, a lightweight cybersecurity model built specifically for threat detection that scored 96% on CyberGym benchmarks.