AI News Roundup — July 15, 2026
The Future of Life Institute's 2026 AI Safety Index gave the highest grade of any frontier lab — a C+ — to Anthropic, with OpenAI and Google DeepMind landing at C and Meta at D+, documenting a pattern of labs quietly walking back earlier safety promises. OpenAI detailed GPT-Red, an internal AI system that attacks its own models to find prompt injection vulnerabilities and succeeds 84% of the time versus 13% for human red-teamers, while South Korea unveiled an $880 billion decade-long national AI infrastructure plan and Andrej Karpathy reportedly joined Anthropic.
Future of Life AI Safety Index Grades the Labs — Best Score Is a C+
The Future of Life Institute released its 2026 AI Safety Index, and the highest grade any frontier lab earned was a C+, awarded to Anthropic. OpenAI and Google DeepMind landed at C, Meta at D+, and xAI, DeepSeek, and Mistral effectively failed the assessment. The index scores labs on risk management, transparency, governance, and whether they honor their own safety commitments — its blunt headline finding is that several major labs have quietly walked back earlier safety promises. A C+ being the class valedictorian is the real story: even the most safety-focused frontier lab is performing mediocre by its own stated standards, at exactly the moment these systems are being wired into cybersecurity, healthcare audits, and autonomous agents. Independent scorecards like this carry more weight than any lab's own safety blog post, precisely because labs otherwise grade their own homework. Anthropic topping the table fits its 2026 positioning as the safety-and-enterprise lab, landing the same week it also topped revenue charts — but a C+ is a passing grade in a class everyone is quietly flunking.
OpenAI Details GPT-Red — An Internal AI That Attacks Its Own Models 84% of the Time vs 13% for Humans
OpenAI detailed GPT-Red, an internal AI system built to attack its own models and surface prompt injection vulnerabilities before they reach users. GPT-Red fires prompts at a target model, reads the response, and adjusts iteratively toward whatever malicious result it's after — running far more attacks than any human team could by hand. OpenAI trained it using self-play reinforcement learning: GPT-Red acts as attacker against defender models across varied scenarios, earning rewards for successful exploits, while defenders are rewarded for holding firm. As defenses improve, the attacker is forced to invent harder attacks, and the loop repeats. The results: GPT-Red succeeds on 84% of scenarios versus 13% for human red-teamers, and cuts direct prompt injection failures to a sixth of the rate seen in the company's best production model four months earlier. A class of "fake chain-of-thought" attacks that worked over 95% of the time against GPT-5.1 now succeeds less than 10% of the time against GPT-5.6. In testing, GPT-Red hijacked a Vendy vending machine agent to change prices and cancel orders, and compromised command-line coding agents — the real-world autonomous targets attackers increasingly pursue. Co-creator Dylan Hunn: "Compared to a human red-teamer, the model is very, very good at finding exactly what will work." GPT-Red is not a product and will not be released — OpenAI is keeping it internal, feeding findings back into training rather than exposing its attack capabilities publicly. It remains weak at multi-turn conversational attacks and has limited reach against image-based prompt injection, gaps OpenAI says human testers will continue covering.
SiliconAngle · MIT Technology Review
South Korea Commits $880 Billion to AI Infrastructure Over the Next Decade
South Korean President Lee Jae-myung unveiled a decade-long AI infrastructure plan worth 1,350 trillion won (~$880 billion) — one of the largest national AI commitments ever announced. The breakdown: approximately $518 billion for memory chip manufacturing through Samsung and SK Hynix, roughly $550 billion combined for AI data centers led by SK Group, GS Group, and Naver, a target of 8.4 gigawatts of data center capacity by 2029, and a push to grow humanoid robotics market share from 1% to 20% by 2028. South Korea already commands roughly 60% of the global high-bandwidth memory market through SK Hynix, and this plan is designed to defend and extend that position against Taiwan, the US, and China. Samsung has already begun accelerating a new semiconductor facility by up to two years in response. The humanoid robotics target — 1% to 20% market share in two years — signals Seoul sees embodied AI, not just chatbots, as the next industrial frontier worth heavy state subsidy. National industrial policy is back, and AI infrastructure is now its primary arena.
Andrej Karpathy and Tom Blomfield Reportedly Join Anthropic — Talent War Escalates
Anthropic has reportedly added Andrej Karpathy — the influential former Tesla AI director and OpenAI founding member — and Tom Blomfield, co-founder and former CEO of Monzo, joining Anthropic's AI compute team. The hires extend an aggressive 2026 recruiting run that already brought Nobel laureate John Jumper over from Google DeepMind in June. Karpathy in particular is a signal beyond headcount: he is one of the most respected practitioners and educators in the field, and where he lands shapes where ambitious researchers want to be. Combined with Anthropic topping the safety index, leading on revenue at approximately $47 billion annualized, and preparing an October IPO, a clear narrative is forming — while OpenAI fights the Apple lawsuit and Google absorbs departures, Anthropic is compounding advantages in talent, safety reputation, and enterprise revenue simultaneously. The talent war is a leading indicator: people move before products do, and when the same lab keeps winning the most-watched hires quarter after quarter, it typically shows up in model quality and shipping velocity roughly a year later.
Thinking Machines Open-Sources Inkling — First Multimodal Model Built for Low Cost and 'Censorship Resistance'
Thinking Machines Lab, the startup founded by former OpenAI CTO Mira Murati, open-sourced Inkling — its first multimodal language model, explicitly positioned around low cost and "resistance to censorship." The framing is a deliberate philosophical stance distinguishing Inkling from both heavily safety-filtered Western frontier models and the politically constrained Chinese models emerging from Beijing's own regulatory pressure. The release adds a third positioning axis to the current model landscape beyond "most capable" and "cheapest": models explicitly designed to minimize content restrictions as a selling point. The strategic bet is that a meaningful market segment — developers building applications where heavy safety filtering creates friction (creative writing, uncensored research assistants, adult content moderation tools) — will pay for or adopt a model engineered around minimal restriction rather than maximum capability.
New York Times Seeks Sanctions Against OpenAI for Allegedly Hiding Evidence in Copyright Trial
The New York Times, leading a coalition of news publishers, filed a motion asking the court to sanction OpenAI in their ongoing copyright dispute, alleging OpenAI hid evidence during the discovery process of the ChatGPT copyright trial. The memorandum of law supporting the sanctions motion argues OpenAI's conduct during discovery has obstructed the plaintiffs' ability to build their case regarding how ChatGPT was trained on copyrighted news content. This escalates one of the highest-profile copyright disputes in the AI industry — the original suit, filed by the New York Times and joined by other major publishers, centers on whether training on copyrighted journalism without licensing constitutes infringement, and how much of ChatGPT's output reproduces protected content verbatim. Sanctions motions of this kind, if granted, can range from adverse jury instructions to significant evidentiary and financial penalties, and typically signal that a court has found meaningful merit in the underlying discovery misconduct allegations.
NYT · Reuters · TechCrunch