A roundup of AI stories that actually mattered on July 20

Trick-Height-3448 · reddit · 2026-07-20

A roundup of AI stories the author thought were genuinely worth reading:

GPT-5.6 and a WordPress exploit chain: A security researcher reportedly used GPT-5.6 Sol Ultra to discover an exploit path from SQL injection to RCE on WordPress. The reported model cost was about $25, while exploit brokers may pay up to $500,000 for a WordPress RCE. WordPress has already shipped fixes.
Claude and a 1939 math conjecture: Claude Fable 5 reportedly produced a hand-checkable counterexample to the Jacobian Conjecture. Mathematicians are now examining it, but the author stresses that the result still needs broader verification.
Kimi K3 vs. safety restrictions: One post claims Kimi K3 fixed 15 critical security bugs in about 10 hours for roughly $250, while OpenAI Codex and Claude Fable 5 refused the task due to cybersecurity safeguards. The author treats this as an interesting but unverified claim and wants the underlying methodology.
OpenAI and Anthropic in U.S. public health: A new program will let public health organizations across 10 U.S. jurisdictions test enterprise AI tools from OpenAI and Anthropic, potentially covering up to 2,000 practitioners. Possible uses include biosurveillance, drug-wave prediction, multilingual communication, operations, and clinical-data retrieval, but privacy, review, retention, and evaluation details are still unclear.
OpenAI free credits: The post also starts discussing whether OpenAI’s recent credit giveaways are as generous as they look, noting concerns about expiration timing and limit resets.

The throughline is that the biggest AI value may now be in security, science, and real operations—not just benchmarks.

More from Safety