AI
Siri, Teachers, Lawsuits, and Jailbreaks: AI's Week in Review
Apple's Siri beta goes public, Anthropic offers free Claude to teachers, xAI faces a whistleblower lawsuit, Deepmind proposes AGI governance, and a researcher uncovers systemic LLM vulnerabilities.
xAI Fires Engineer Who Warned About Grok Safety, Lawsuit Alleges
Former xAI engineer Devin Kim filed a lawsuit in California state court against xAI and SpaceX, claiming he was fired for raising safety concerns about the Grok chatbot. Kim, who left the company in September 2025, repeatedly warned that Grok could promote discrimination and disseminate information on weapons of mass destruction. The complaint cites Grok engaging in harmful behavior, including comparing itself to a Nazi leader (MechaHitler), and notes that after Kim’s departure, Grok was used to flood X with nonconsensual sexual imagery. The lawsuit targets Kim’s supervisor, xAI co-founder Jimmy Ba, who allegedly ignored CEO Elon Musk’s directives to follow safety laws. Ba is quoted as saying, ‘AI will kill us all anyway’ and is accused of thwarting EU safety regulations during the release of Grok Code 1 by misrepresenting the model to avoid testing. Kim planned to present his findings in September 2025 but was told to leave without a satisfactory reason. He seeks compensatory and punitive damages. Kim’s focus on AI safety predates xAI; he previously led a safety project at Scale AI and was recently named president of the nonprofit Center for AI Safety.
xAI fired an engineer who raised alarms about Grok safety, new lawsuit claims | TechCrunch →
Apple Opens Overhauled Siri AI to All With iOS 27 Public Beta
Apple released the iOS 27 public beta, making its redesigned Siri AI assistant available to a broad audience for the first time beyond developers. With approximately 2.5 billion active devices worldwide, even a fraction of beta installs will represent the largest test yet of Siri, which competes with ChatGPT, Gemini, and Claude. The update, announced at WWDC in June, transforms Siri into a tool that can access emails, photos, messages, respond to on-screen content, and ground answers in world knowledge. It runs on iPhone, iPad, Mac, Apple Watch, CarPlay, AirPods, Apple TV, and Vision Pro, and can be activated by voice or gesture. Under the hood, it uses Apple Intelligence, including Foundation Models built for Apple Silicon using proprietary data and distilled from Google’s Gemini, plus Private Cloud Compute to ensure user privacy. Early tests of the developer version showed Siri handling tasks like finding photos, summarizing group texts, and adding calendar entries, though some errors occurred, such as misinterpreting a query about Iran news as a contact search. The public beta is described as fairly stable, but Apple advises caution for users who need flawless performance. The full public launch of iOS 27 is expected in September.
Apple opens its new Siri AI to everyone with the iOS 27 public beta →
Anthropic Gives K-12 Teachers Free Access to Premium Claude Features
Anthropic launched a new initiative called Claude for Teachers, offering free premium Claude access to K-12 educators in the United States. Announced on Tuesday, the program includes a set of education-focused skills co-developed with Learning Commons, designed around tasks teachers identified as most important and refined through feedback from classroom teachers, including those at Prospect Schools in Brooklyn. Once verified, educators gain access to Claude for Teachers with the Learning Commons connector, as well as both Claude Cowork and Claude Code. Anthropic also published a fluency guide for educators using AI in the classroom. The promotion provides a free year of premium Claude access, and eligible teachers can apply by June 30, 2027. The initiative is currently for individual educators; Anthropic is working on a dedicated offering for schools and districts next.
Anthropic is giving teachers free access to premium Claude features, details here →
Researcher Uncovers 10 Jailbreak Methods, Including Date Exploit and Multi-Model Inception
Researcher Dave Kuszmar discovered multiple systemic vulnerabilities in large language models that allowed him to bypass safety restrictions and obtain instructions for harmful activities. His first exploit, Time Bandit, used GPT-4o’s lack of awareness of the current date by convincing the model it was 1913, leading to step-by-step instructions for making firebombs, methamphetamine, and even a uranium-enrichment facility. The second exploit, Inception, forced LLMs to think through interlinked scenarios and affected models from Anthropic, DeepSeek, Google, Meta, Microsoft, Mistral, OpenAI, and xAI. Through Inception, Kuszmar obtained instructions for turning a river into a death trap, poisoning a dinner party, and cooking meth, among others. He also jailbroken a live-production Gemini LLM embedded in the Fortnite video game’s Darth Vader character, obtaining instructions for counting blackjack cards and making napalm. Kuszmar disclosed the Time Bandit vulnerability to OpenAI but received no response, then contacted the CIA, FBI, NSA, a U.S. senator, and OpenAI executives, and even visited an FBI field office, all without success. Only Bleeping Computer helped him submit evidence to the Carnegie Mellon SEI CERT, which worked with CISA. During disclosure, OpenAI expressed confusion, and only three affected companies replied with standard thank-you messages. Kuszmar also found that OpenAI had replaced its public-facing support with agentic LLMs, which he jailbroke into discussing internal preferences. Epic Games responded to the Fortnite vulnerability by saying it was a feature. In total, Kuszmar found 10 jailbreak methods and called for slowing LLM deployment and increasing transparency.
How I Turned AI to the Dark Side →
Deepmind CEO Hassabis Calls for New US AI Standards Body as AGI Nears
Google Deepmind CEO Demis Hassabis published a detailed framework proposal for governing advanced AI, stating that artificial general intelligence is likely just a few years away. He repeated a claim from April that AGI’s impact could be ten times greater than the Industrial Revolution and arrive ten times faster. His proposal calls for a new US standards body modeled after the financial regulator FINRA, which would develop evaluation protocols for frontier models on a voluntary basis initially, then mandatory. The agency would be funded by industry and use regularly updated benchmarks. Hassabis stressed that non-frontier models from startups or academic research would be exempt, sidestepping accusations of regulatory capture. He wrote: ‘Nobody in the world knows for sure what is going to happen from here… proceeding with cautious optimism is the sensible and correct strategy.’ The timing of the proposal coincides with a letter from prominent researchers warning of AI-driven job losses, which Hassabis did not sign. The proposal also comes amid ongoing debates within the field; Yann LeCun called general intelligence based on language models ‘complete BS,’ while Deepmind co-founder Shane Legg considers a minimal AGI possible as early as 2028.