8 May 2026: AlphaEvolve’s Real-World Results, Firefox’s AI Security Leap and OpenAI’s Voice Overhaul (AM)

Voice AI reached a new level of sophistication overnight, Anthropic’s security research arm exposed 271 vulnerabilities in Firefox, and Google detailed what its algorithm-discovery AI has been quietly achieving across science and infrastructure. This morning’s AI news shows the technology delivering measurable results, not just promises.

Google DeepMind has published a detailed impact report on AlphaEvolve, its Gemini-powered algorithm discovery agent, with results that stretch from genomics to chip design. In DNA sequencing, AlphaEvolve improved DeepConsensus, a model that corrects errors in genetic sequencing data, cutting variant detection mistakes by 30 percent. The system has also been used to optimise algorithms running across parts of Google’s own infrastructure, with efficiency gains that touch millions of users daily.

AlphaEvolve works by using Gemini’s language model capabilities to propose, test, and refine algorithms automatically, a task that previously required months of specialist engineering effort. DeepMind says the system has exceeded all previously known solutions on several open mathematical problems, and that results on chip design and matrix multiplication are already in production use at Google. Access is not yet public, but DeepMind has signalled plans to expand availability over the coming months.

Anthropic’s Claude Mythos Preview has transformed how Mozilla handles security testing for Firefox: in April 2026, Firefox shipped 423 bug fixes, compared with just 31 in the same month a year earlier. The jump is almost entirely attributable to Mythos, which found 271 security vulnerabilities including a 15-year-old flaw in the HTML legend element, a 20-year-old XSLT bug, and memory safety issues that could have allowed a compromised process to manipulate the browser from within.

Mythos is Anthropic’s agentic system built specifically for security research. Its deployment at Mozilla is one of the clearest examples yet of AI delivering measurable safety improvements to consumer software used by hundreds of millions of people. More than 40 CVEs (Common Vulnerabilities and Exposures, each a confirmed distinct security flaw) were patched across Firefox 149 and 150 as a direct result.

Mozilla notes that all the flaws were theoretically findable by an elite human researcher. The practical difference is throughput: Mythos processed millions of lines of code in the time it would take a human team weeks to review. For anyone using Firefox, versions 149.0.2 and 150 are now meaningfully more secure.

OpenAI has released a new generation of realtime voice models through its API, pushing well beyond simple speech-to-text transcription. The new models can reason about what they hear, translate speech across languages, and generate conversational responses, all within a single low-latency loop. This is a significant advance on the previous setup, which handled transcription and response generation as separate tasks.

The applications are broad. Customer service platforms can now build systems that understand context and intent in spoken conversation rather than relying on scripted menus. Education tools can support real-time tutoring in multiple languages. Parloa, an enterprise voice agent platform, is already building on the new models for large-scale customer service deployment. UK developers can access the new voice API through OpenAI’s standard platform, with pricing for reasoning and translation features not yet published separately.

OpenAI has added a Trusted Contact safeguard to ChatGPT, letting users designate someone who will be alerted if their conversations show signs of self-harm risk. The feature is entirely opt-in: users set it up in advance, naming a friend, family member, or mental health professional. When the system detects concern in a conversation, it surfaces helpline information to the user and simultaneously notifies the registered contact.

The addition reflects a growing sense of responsibility among AI companies toward vulnerable users. The Trusted Contact system attempts to link ChatGPT’s automated safety responses to real-world support networks rather than relying solely on chatbot-level interventions. UK users who configure the feature will see NHS crisis support information surfaced alongside any notification sent to their contact. OpenAI says the rollout is global.

Perplexity has opened its Personal Computer product to all Mac users, ending the limited beta that had restricted access since early 2026. Personal Computer is an AI agent that runs natively on your Mac and can take actions across applications: searching the web, summarising documents, drafting messages, and navigating software on your behalf. Unlike browser-based AI tools, it has direct access to your desktop environment.

This expansion makes Perplexity one of the first major AI companies to offer a fully open desktop agent to consumers. For small business owners who spend significant time switching between applications or answering routine queries, it is designed to absorb that overhead. Basic use is free, with a Pro tier for more advanced tasks.

Here is everything else worth knowing from this morning’s AI news.

• GPT-5.5 and GPT-5.5-Cyber available to verified security defenders [7 May]: OpenAI is expanding Trusted Access for Cyber, giving verified professionals access to its latest models to accelerate vulnerability research and protect critical infrastructure.
• Bumble is removing the swipe mechanic [7 May]: The dating app is pivoting toward AI-assisted matching, including an AI assistant called Bee that learns user preferences over time and suggests compatible matches.
• Stockholm AI startup Pit raises a $16 million seed round from a16z [7 May]: Founded by the co-founders of European scooter company Voi, Pit is one of Europe’s fastest-rising AI startups with Andreessen Horowitz leading.
• Elon Musk’s lawsuit against OpenAI focuses on its founding mission [7 May]: Legal arguments centre on whether OpenAI’s for-profit structure is compatible with its original charter to ensure AI benefits humanity broadly, not just shareholders.
• Anthropic published research on Natural Language Autoencoders [7 May]: The technique represents Claude’s internal reasoning as human-readable text, a step toward making AI decision-making more interpretable and auditable.
• AI-generated content is degrading online communities [7 May]: Authentic participation in forums is falling as automated posts increase, with community managers reporting a sharp drop in genuine exchange.
• Parloa is deploying OpenAI’s new voice models for enterprise customer service [7 May]: The voice AI platform is among the first building on the new realtime reasoning API, targeting large-scale inbound customer service for major businesses.

This is a daily news update for informational purposes only. AI products and policies change rapidly. Verify details directly with providers before making decisions. Nothing here is financial or legal advice.

AI Daily is Cristoniq’s daily guide to developments in artificial intelligence, published every morning.