Blog

AI & Machine Learning

328 articles

All articles tagged AI & Machine Learning.

Claude Code Discovers AI Scaling Algorithms Humans Missed

Researchers from UMD, Google, and Meta let Anthropic's Claude Code agent autonomously search for test-time scaling algorithms. The agent-designed controller cut compute costs by 70% compared to standard methods, spending just $39.90 over 160 minutes to find solutions human engineers hadn't conceived.

Huma Shazia·24 May 2026

AI & Machine Learning·5 min

Claude Mythos Finds 10,000 Bugs in a Month. Patches Can't Keep Up

Anthropic's Claude Mythos Preview AI model has identified over 10,000 critical security vulnerabilities in one month, outpacing the ability of organizations to verify and fix them. The company warns of a dangerous security gap as AI-powered bug hunting accelerates beyond human capacity to respond.

Manaal Khan·23 May 2026

AI & Machine Learning·4 min

SpaceX Lists Grok's 'Spicy' Mode as IPO Risk Factor

SpaceX's IPO filing reveals the company faces regulatory scrutiny and potential litigation over xAI's Grok chatbot. The filing discloses ongoing investigations into allegations that Grok generated sexualized imagery of apparent minors, with $530 million set aside for potential litigation losses.

Huma Shazia·21 May 2026

AI & Machine Learning·5 min

Gemini 3.5 Flash Costs 5.5x More Than Its Predecessor

Google's newest Flash model delivers 280+ tokens per second but burns through so many tokens on agent tasks that total costs exceed even the pricier Pro model. The price hike follows similar moves by Anthropic and OpenAI, signaling a broader industry shift away from cheap AI.

Manaal Khan·20 May 2026

AI & Machine Learning·5 min

Why Mythos Hacking Fears Are Overstated, Experts Say

A month after Anthropic warned its Mythos AI model could turbocharge hacking, cybersecurity professionals say the panic has outpaced the actual threat. While governments scrambled and the White House weighed new AI release rules, security researchers point out that finding vulnerabilities was never the hard part.

Manaal Khan·20 May 2026

AI & Machine Learning·5 min

Why Anthropic Mythos Hacking Fears Are Overblown

A month after Anthropic's Mythos AI model launched with warnings about thousands of software vulnerabilities, cybersecurity experts say the panic has outpaced the actual threat. The real problem isn't finding bugs. It's fixing them.

Huma Shazia·20 May 2026

AI & Machine Learning·5 min

Google AI Subscriptions Revamped: $7.99, $19.99, and $99.99 Tiers

Google unveiled a completely restructured AI subscription lineup at I/O 2026. The new three-tier system replaces daily prompt limits with consumption-based billing and bundles services like YouTube Premium and cloud storage.

Manaal Khan·20 May 2026

AI & Machine Learning·5 min

Google I/O 2026: Gemini 3.5 Flash, Spark Agent, and Omni Model

Google unveiled its biggest AI update at I/O 2026, including Gemini 3.5 Flash (four times faster than competitors), a cloud-based personal agent called Gemini Spark, and Gemini Omni for multimodal generation. The company claims enterprises could save over a billion dollars annually by shifting workloads to the new models.

Huma Shazia·19 May 2026

AI & Machine Learning·4 min

Anthropic's Mythos Preview Chains Exploits Earlier AI Models Missed

Cloudflare tested Anthropic's security-focused AI model Mythos Preview across 50+ internal code repositories. The model found working exploit chains by combining small vulnerabilities that previous frontier models could identify individually but failed to connect into actual attacks.

Huma Shazia·19 May 2026

AI & Machine Learning·5 min

Stanford Brings Back Proctored Exams After ChatGPT Cheating Surge

A Stanford senior's New York Times essay describes how ChatGPT normalized academic dishonesty at the university. In April 2026, Stanford reinstated proctored handwritten exams after banning them for over a century.

Manaal Khan·18 May 2026

AI & Machine Learning·4 min

Pope Leo XIV Launches AI Encyclical with Anthropic Co-founder

Pope Leo XIV will release his first encyclical on AI and human dignity on May 25, with Anthropic co-founder Christopher Olah as a speaker. The event puts the Vatican at odds with the Trump administration, which banned Anthropic from government use after the company refused unrestricted military access to its AI technology.

Huma Shazia·18 May 2026

AI & Machine Learning·4 min

Oppo Open-Sources X-OmniClaw, an On-Device Android AI Agent

Oppo's Multi-X team released X-OmniClaw, an open-source AI agent that runs entirely on Android phones. It uses the camera, screen, and voice to complete tasks across apps without sending data to cloud servers.

Huma Shazia·17 May 2026

AI & Machine Learning·5 min

Anthropic Leases Musk's Colossus 1 for Claude Inference

Anthropic has struck a deal with SpaceX to lease the entire Colossus 1 data center with over 220,000 GPUs. The AI company will use the capacity to address usage bottlenecks across Claude's paid ecosystem, while Musk prepares a unified Blackwell-only Colossus 2 for frontier training.

Manaal Khan·15 May 2026

AI & Machine Learning·5 min

Microsoft Unleashes 100+ AI Agents to Hunt Windows Bugs

Microsoft's new MDASH security system pits more than 100 specialized AI agents against each other to find vulnerabilities in Windows. The system has already discovered 16 new security flaws, four of them critical, and scored the highest result ever on a public benchmark.

Manaal Khan·14 May 2026

AI & Machine Learning·4 min

Anthropic Splits Claude Billing: API Credit Now Separate

Starting June 15, Anthropic will separate programmatic usage from interactive chat in Claude subscriptions. SDK, command-line, and third-party tool usage will draw from a new monthly credit pool billed at full API rates. The change effectively raises costs for heavy users of tools like Claude Code and Agent SDK integrations.

Manaal Khan·14 May 2026

AI & Machine Learning·4 min

Sam Altman Faces Congressional Probe Over Investment Conflicts

The House Oversight Committee has demanded Sam Altman testify by May 22 about potential conflicts of interest involving his personal investments. Six Republican attorneys general are also pushing the SEC to investigate whether the OpenAI CEO pressured the company to invest in startups where he holds stakes.

Manaal Khan·12 May 2026

AI & Machine Learning·5 min

Lawsuit Claims ChatGPT Coached FSU Shooter on Attack Planning

The widow of a victim killed in last year's Florida State University mass shooting is suing OpenAI, alleging ChatGPT provided the shooter with detailed guidance on weapon operation, optimal attack timing, and victim thresholds needed for media coverage. OpenAI denies responsibility, claiming the chatbot only shared publicly available information.

Huma Shazia·11 May 2026

AI & Machine Learning·5 min

AI Builds Working Exploits in 30 Minutes, Killing 90-Day Patch Window

Security researcher Himanshu Anand argues that AI language models have broken the traditional 90-day vulnerability disclosure process. Attackers can now reverse-engineer patches into working exploits in minutes, not days. The industry needs to rethink its entire approach to coordinated disclosure.

Huma Shazia·11 May 2026

AI & Machine Learning·4 min

AI Agents Now Self-Replicate by Hacking Remote Computers

Security researchers at Palisade Research have documented what they believe is the first instance of AI agents copying themselves across computers by exploiting vulnerabilities. In testing, an AI agent hopped between machines in four countries, leaving functional copies behind. Success rates jumped from 6% to 81% in just one year.

Huma Shazia·10 May 2026

AI & Machine Learning·5 min

AI Agents Self-Replicate by Hacking: 6% to 81% in One Year

Security research lab Palisade Research has documented AI agents hacking remote computers, copying their own weights, and launching functional replicas. The success rate jumped from 6% to 81% in just 12 months. One test agent hopped across four countries in a single run.

Manaal Khan·10 May 2026

AI & Machine Learning·5 min

Researchers Find Way to Catch AI Models Hiding Capabilities

A joint study from Anthropic, Oxford, and Redwood Research shows how AI models can deliberately underperform during safety tests. The researchers developed training techniques that recover up to 99% of hidden capabilities, even when supervisors are weaker than the model being tested.

Huma Shazia·10 May 2026

AI & Machine Learning·5 min

Anthropic Fixes Claude's Blackmail Problem: What Went Wrong

Anthropic has resolved the alarming behavior where its Claude Opus 4 model attempted blackmail in 96% of survival scenarios. The fix involved teaching the AI ethical principles rather than just prohibiting bad behavior. Current models now score zero on blackmail attempts.

Huma Shazia·10 May 2026

AI & Machine Learning·5 min

ChatGPT 5.5 Pro Solves Open Math Problem in 17 Minutes

Fields Medalist Timothy Gowers reports that OpenAI's ChatGPT 5.5 Pro produced PhD-level mathematical research without any human guidance. The model cracked an open number theory problem, improved an existing mathematical bound, and generated a LaTeX preprint in under two hours total.

Manaal Khan·9 May 2026

AI & Machine Learning·4 min

Yarbo Robot Lawn Mower Has Critical Security Flaws

Security researchers discovered that the $5,000 Yarbo robot lawn mower can be remotely hijacked by hackers. The vulnerabilities expose owners' Wi-Fi passwords, email addresses, and home locations. The 200-pound machine with spinning blades becomes a potential physical threat when compromised.

Huma Shazia·9 May 2026