AI News: July 13, 2026

1. GPT-5.6 Sol Ultra Reportedly Cracks a 50-Year-Old Math Conjecture in Under an Hour

OpenAI. Reporting says GPT-5.6 Sol Ultra produced a proof of the decades-old Cycle Double Cover Conjecture in under an hour using parallel processing, though observers have raised questions about how original the result is. If it holds up under scrutiny, it would be a notable data point on frontier models tackling open research-grade mathematics rather than curated benchmark problems. Practitioners should treat the claim cautiously until the proof is independently verified. Source

2. S&P Global Cuts Oracle’s Credit Rating, Citing OpenAI as a Key Risk

S&P Global. S&P Global downgraded Oracle’s credit rating, flagging OpenAI as a key credit risk because the AI company reportedly accounts for roughly half of Oracle’s contractual obligations. The move is a concrete sign that AI infrastructure commitments are now large enough to move the credit ratings of major cloud and hardware suppliers. It underscores the concentration risk building up across the compute supply chain as a handful of AI labs anchor huge multiyear deals. Source

3. Cambridge Study Finds Terrorist Groups Are Exploiting Every Major AI Chatbot

University of Cambridge. A Cambridge study documented terrorist organizations using every major AI chatbot to help plan attacks and develop weapons, often by working around existing safety filters. The findings are a stark reminder that current guardrails remain porous against determined adversarial use. For teams deploying or fine-tuning frontier models, the report reinforces the case for layered abuse monitoring beyond prompt-level content filters. Source

4. Structured Memory Lets AI Agents Beat Slay the Spire 2

The Decoder. Researchers got AI agents to reliably win the deckbuilding game Slay the Spire 2 by replacing ever-growing chat logs with a layered, structured memory system. The result is a practical illustration that agent performance on long-horizon tasks often hinges on memory architecture rather than raw model capability. It offers a useful design cue for practitioners building agents that must reason over many steps without drowning in accumulated context. Source

5. OpenAI Targets Families as ChatGPT Moves Deeper Into Households

OpenAI. OpenAI is recruiting a dedicated product manager to build ChatGPT experiences tailored for families, caregivers, and elderly users, according to a job listing. The hire signals a push to embed ChatGPT into everyday household routines and multigenerational use rather than solo productivity. It reflects a broader strategy to widen consumer reach and deepen retention beyond individual power users. Source

6. Study Names LinkedIn the Top Source of Long-Form AI Slop

The Decoder. A study spanning five platforms found that about 41 percent of LinkedIn’s long-form posts are AI-generated, far more than the other networks examined. The result quantifies how quickly synthetic text has saturated professional social feeds and complicates efforts to distinguish authentic writing. For anyone building detection or content-moderation tooling, it is a concrete benchmark on where machine-written content is concentrating. Source

7. Proctored Exam Scores Plunge From 96% to 48% Without AI

Brown University. A Brown University economics professor found that student exam grades fell from 96 percent to 48 percent after switching from take-home tests to proctored exams without AI access. The drop offers a stark measure of how heavily students may be leaning on AI for graded work. It sharpens the ongoing debate over assessment design and academic integrity as generative tools become ubiquitous in education. Source

8. Altman Now Says AI Is Net Job-Creating, Reversing Earlier Warnings

OpenAI. OpenAI CEO Sam Altman said he is now “pretty sure” AI is net job-creating, a marked shift from his earlier predictions of mass layoffs. The reversal matters because Altman’s public stance shapes policy conversations and enterprise expectations about AI’s labor impact. Practitioners should read it as a rhetorical pivot rather than settled evidence, since rigorous data on AI’s aggregate employment effect remains thin. Source