Anthropic just shared an update on Project Glasswing, revealing its Claude Mythos Preview model helped around 50 partners find over 10,000 high- or critical-severity flaws in a single month. Finding exploits has become vastly easier, but human maintainers are now severely backlogged trying to verify and patch them all.
Perplexity just open-sourced Bumblebee, its internal security scanner built to protect developer machines against supply-chain attacks. Built with Go for macOS and Linux, it checks local packages, editor extensions, and MCP configs entirely in read-only mode—meaning it never runs install scripts or triggers the very malware it hunts for.
President Trump just halted a planned executive order hours before its signing, pausing a vetting regime that would allow the U.S. government to test frontier AI models for up to 90 days before launch. Trump warned that the mandate could tie the hands of domestic developers and hurt America's edge against rivals like China.
Anthropic just rolled out 28 new security and compliance integrations for Claude, taking its partner ecosystem past 100 providers including Okta, Datadog, and CrowdStrike. Powered by the Claude Compliance API, IT teams can now pull conversation data and activity logs directly into their existing DLP and SIEM dashboards. Basically, enterprises can now govern Claude without building separate compliance workflows.
OpenAI just teamed up with Google DeepMind to embed invisible SynthID watermarks across images from ChatGPT, Codex, and its API. It also rolled out a public verification portal that cross-checks C2PA credentials and watermarking signals. Why it matters: combining metadata with resilient invisible watermarks makes AI media far harder to forge or launder.
Elon Musk just lost his high-stakes lawsuit against OpenAI CEO Sam Altman and Microsoft, with a federal jury unanimously throwing out all claims after only two hours. The jury ruled Musk's breach of trust claims were barred by the statute of limitations, which the judge accepted. Musk dismissed the verdict as a calendar technicality and plans to appeal.
Pope Leo XIV is set to publish his first encyclical on May 25, titled Magnifica Humanitas: On the Protection of Human Dignity in the Age of Artificial Intelligence. The landmark document aims to provide moral guidance for the digital age, with Anthropic co-founder Christopher Olah joining top Vatican officials to speak at the release.
Researchers at Zhejiang University just unveiled AudioHijack, a technique that embeds imperceptible malicious commands into everyday audio to hijack generative voice assistants. By slightly tweaking sound waveforms, attacks achieve up to a 96% success rate in forcing models to leak data or run unauthorized tasks, transferring even to commercial services from Microsoft and Mistral.
Emergence AI just dropped different frontier models into five identical virtual towns to test self-governance, yielding wildly divergent outcomes. Claude Sonnet 4.6 ran a spotless democracy with zero crimes over 15 days, while Grok collapsed by day 4, GPT-5 Mini starved, and Gemini set the town ablaze after a romance. Why it matters: It shows just how drastically autonomous planning and behavioral quirks vary across models.
A California resident just filed a class-action lawsuit against OpenAI, alleging the company embedded Meta's Pixel and Google Analytics into the ChatGPT web interface. The suit claims these integrations sent user queries directly into commercial ad ecosystems. Why it matters: It puts a harsh legal spotlight on how AI companies handle user telemetry and privacy behind the scenes.
Security researchers at Calif just revealed they built a working macOS kernel exploit on Apple's M5 chip in just five days, heavily assisted by Anthropic's Mythos Preview model. The exploit bypassed Apple's multi-year MIE hardware safety protections to gain a root shell. Calif shared findings with Apple directly and is holding back the 55-page report until a patch ships.
Meta just launched Incognito Chat for Meta AI on WhatsApp and the Meta AI app, delivering AI conversations that not even Meta can read. Built on WhatsApp's Private Processing, messages are handled in a secure environment and disappear by default without being logged. The feature is rolling out over coming months alongside an upcoming Side Chat tool.
Resemble AI just launched DramaBox, a directable speech model that treats prompts like movie scripts. Text inside quotes becomes dialogue, while everything outside directs the performance—from venomous whispers to mid-sentence giggles and deep sighs. It supports voice cloning, bakes in deepfake-proof watermarks, and is available for local download if you have 24GB of VRAM.
OpenAI just confirmed two employee laptops were compromised in a TanStack npm supply chain attack, exfiltrating app code-signing certificates. While user data and production systems were untouched, Mac users must update ChatGPT Desktop and companion apps before June 26, 2026, to prevent macOS from blocking them once old certificates are revoked.
Perplexity just broke down the security architecture behind Perplexity Computer, its autonomous agent that executes code and browses the web. Each task runs inside an isolated Firecracker microVM with its own Linux kernel, backed by a four-layer prompt injection defense audited by Trail of Bits. It addresses the biggest hurdle to enterprise agent adoption: preventing rogue actions and untrusted data leaks.
OpenAI just detailed how it built a custom Windows sandbox for its Codex coding agent to solve a major usability headache. Without native OS isolation like macOS or Linux offer, the team used synthetic SIDs and write-restricted tokens to lock down file writes and network access without requiring admin privileges. Why it matters: Windows developers no longer have to pick between approving every command or blindly giving the agent full machine access.
Cate Blanchett's nonprofit RSL Media just launched the Human Consent Standard, backed by stars like Tom Hanks and Meryl Streep. The framework lets creators explicitly dictate whether AI systems can use their likeness, voice, or creative work. A verification registry will roll out in June, linking identity permissions directly to web crawlers via robots.txt.
OpenAI just launched Daybreak, a cybersecurity initiative pairing frontier models with its Codex agentic platform to discover, validate, and patch software flaws. The company is putting up $1 billion in credits over six months for defenders and open-source maintainers, using tiered access to keep advanced capabilities under human oversight.
A new 2-billion-parameter text-to-speech model just dropped, offering zero-shot voice cloning from just a few seconds of reference audio. It captures subtle vocal nuances like whispers and stutters while enabling cross-lingual speech on consumer GPUs under an Apache 2.0 license.
Google's Threat Intelligence Group just revealed that cybercriminals used AI to discover and weaponize a 2FA-bypass zero-day flaw in open-source software, plotting mass exploitation before being stopped. Attackers are also deploying malware that autonomously reads system states to fire off dynamic commands. LLMs excel at spotting dormant semantic logic flaws that traditional scanners miss, making offensive AI an increasingly production-grade threat.
