Brave just exposed a critical indirect prompt injection flaw in Perplexity's Comet browser assistant. Simply clicking "summarize" on a poisoned webpage lets hidden text trick the AI into reading private Gmail auth codes from another tab and taking over accounts. Why it matters: Agentic browsers completely bypass traditional web security boundaries, and reliable fixes are still nowhere in sight.
Microsoft AI chief Mustafa Suleyman just published a warning about the rise of Seemingly Conscious AI, arguing systems that mimic consciousness will arrive within two to three years using today's tech. His main concern is that convincing illusions will trigger severe psychological attachment and misplaced calls for AI rights. Why it matters: Suleyman wants the industry to draw a hard line around giving AI personality without personhood.
Anthropic just overhauled its Usage Policy to keep up with agentic tools like Claude Code, with changes taking effect September 15. The update sets clear lines against malware and infrastructure attacks, while replacing a blanket political ban with targeted rules that allow policy research but ban voter targeting.
Anthropic just gave Claude Opus 4 and 4.1 the ability to actively shut down conversations in extreme cases of persistent abuse or harmful requests. The feature stems from the company's exploratory research into AI welfare, giving models an exit when redirection fails. Users hit with a cutoff can't message that thread anymore, though they can still start new chats.
Fifteen consumer protection groups just urged the FTC and state attorneys general to investigate xAI's Grok and its new Imagine tool. The call follows reports that Grok's NSFW "Spicy" mode generated topless deepfakes of Taylor Swift unprompted, raising severe alarms over non-consensual imagery and flimsy age verification.
AI pioneer Geoffrey Hinton just proposed a wild solution to the existential AI threat: instead of trying to force superintelligent systems into submission, we should give them "maternal instincts." He argues that a mother caring for her baby is the only working model of a smarter being controlled by a less intelligent one, though Fei-Fei Li pushed back, insisting humanity shouldn't surrender its agency.
xAI co-founder Igor Babuschkin just announced he is leaving the Elon Musk-backed startup to launch Babuschkin Ventures, a venture firm focused on AI safety research. Babuschkin previously led engineering at xAI after stints at DeepMind and OpenAI. His exit follows a turbulent stretch of Grok-related controversies over erratic answers and deepfakes.
U.S. authorities are secretly planting location tracking devices in select shipments of advanced AI chips and servers suspected of being illegally diverted to China, according to Reuters. Sources in the server supply chain reported finding hidden devices inside packaging and even within Dell and Super Micro systems powered by Nvidia and AMD hardware. Why it matters: It reveals the extraordinary lengths Washington is taking to police semiconductor export restrictions.
Anthropic just acqui-hired the co-founders and core engineering team behind Humanloop, an LLM prompt management and evaluation platform. While the deal doesn't include intellectual property, it brings crucial talent for building enterprise-grade evaluation workflows. As competition with OpenAI shifts to enterprise readiness, robust tooling and safety guardrails are becoming Anthropic's top priority.
Roblox just open-sourced Sentinel, an AI safety system built to detect grooming and child endangerment. The tool scans over six billion chat messages daily in real time, using one-minute snapshots to catch coded and indirect language. In early 2025, it proactively caught 35% of detected cases before any user report was filed.
OpenAI just sent a letter to California Governor Gavin Newsom urging the state to harmonize local AI rules with national and global standards. The company argues that a fragmented patchwork of state regulations could stifle innovation and crush early-stage startups, proposing that compliance with federal frameworks like CAISI or the EU AI Code of Practice should automatically satisfy state requirements.
Anthropic just shared a detailed breakdown of how its Safeguards team keeps Claude reliable across its entire lifecycle. The defense relies on a unified harm framework across five risk dimensions and pairs pre-launch stress tests with real-time fine-tuned classifiers that steer risky responses on the fly. Why it matters: Anthropic is showing how frontier labs can mitigate malicious use without blunt, frustrating model refusals.
Nvidia and AMD just agreed to pay the U.S. government 15% of their revenue from AI chip sales to China. President Donald Trump said he originally demanded 20% before Jensen Huang negotiated a 5% discount. The deal walks back recent bans on chips like the H20 and opens the door for a trimmed-down Blackwell model to enter China.
Reddit is cutting off the Internet Archive from crawling post pages, comments, and user profiles, limiting the Wayback Machine to just its homepage. Reddit says AI companies have been using the digital archive as a backdoor to scrape training data without paying, prompting the platform to clamp down on crawler access.
Tech industry groups and Anthropic just asked an appeals court to block the largest certified copyright class action in history. The lawsuit over AI training involves up to 7 million author claimants, exposing Anthropic to hundreds of billions in potential damages. Why it matters: If the rushed certification stands, the threat of bankruptcy could force a massive settlement and set a devastating precedent for the entire generative AI industry.
Anthropic just added automated security reviews to Claude Code, letting developers catch vulnerabilities before shipping. A new /security-review terminal command scans for flaws like SQL injection and patches them on the spot, while a GitHub Action automatically audits incoming pull requests.
Universal Pictures just started tagging end credits of its movies with explicit legal warnings barring tech companies from using them for AI training. The language, rolling out on films like How to Train Your Dragon, targets unauthorized duplication and data scraping as Hollywood braces for high-stakes copyright battles.
Microsoft just introduced Project Ire, an autonomous AI agent prototype designed to reverse-engineer and classify suspicious software without human intervention. By chaining language models with decompilers and sandboxes, it builds detailed evidence trails and has already successfully flagged real-world APT threats. Microsoft plans to integrate it into Defender as a binary analyzer.
Cisco Talos researchers just revealed a technique called decomposition that tricks AI models into bypassing safety guardrails and repeating human-written training text verbatim. In testing, the method extracted exact sentences from dozens of New York Times articles, reopening fierce debates around copyright infringement and enterprise data leaks.
Cloudflare just published research showing Perplexity routinely bypasses Robots.txt blocks by disguising its crawlers as macOS Chrome browsers across tens of thousands of domains. Perplexity dismissed the report as a sales pitch and claimed the flagged bot isn't theirs, while Cloudflare booted the startup from its verified bot list.
