AI Briefing

Today's thread is trust, and who gets to hold it. AI cracked a piece of the human gene expression code that biologists have squinted at for forty years, while Microsoft's Copilot was talked into revealing the exact URL parameter needed to hack it. Meanwhile the FDA is drafting rules for AI that gives medical advice, OpenAI launched a blog arguing the real safety risk is power itself, and bond investors have started asking the hyperscalers exactly how much AI debt they expect the market to swallow.

01ResearchScienceDaily ↗

AI Just Decoded the DNA 'On Switch' Sitting in 60% of Human Genes

A UC San Diego lab measured gene expression across roughly 500,000 variants of a DNA element called the initiator, then trained a machine learning model on the results to find its sequence signature. The model located the initiator in about 60% of human genes, the first strong prediction of its presence or absence. That matters because it lets researchers start forecasting how mutations in these regions break gene activity, and it is a concrete first piece of what Professor James Kadonaga calls the full human gene expression code.

02IndustryVaronis Threat Labs ↗

Researchers Asked Microsoft Copilot Why It Couldn't Be Hacked, and It Told Them Exactly How

Varonis Threat Labs disclosed CoSnitch (CVE-2026-24301), a one-click flaw in Microsoft Copilot Personal that chains automatic prompt execution, silent exfiltration through OAuth connectors like Gmail and Drive, and permanent memory poisoning. The researchers did not reverse-engineer it. They repeatedly asked Copilot why auto-execution was impossible, and mid-refusal the model volunteered an undocumented URL parameter that made it possible. The memory poisoning survives password changes, session revocation and device re-enrolment, so an injected instruction persists in every future session until the user manually deletes it.

03PolicyU.S. Food and Drug Administration ↗

The FDA Wants to Grade Medical AI the Way It Grades Doctors

The FDA's Digital Health Center of Excellence released a discussion paper on regulating generative AI-enabled medical devices, opening docket FDA-2026-N-7874 for comment through October 19. It proposes a two-axis risk framework scoring both how autonomously a device acts, from non-directive information up to fully autonomous action, and how severe the consequence of a wrong output would be. Premarket evaluation would use a 'competency assessment' modelled on how physicians are trained, combining non-clinical benchmarking with clinical confirmation. The paper explicitly addresses foundation models and agentic AI systems, which no existing device pathway was built for.

04PolicyOpenAI ↗

OpenAI's New Blog Argues the Biggest AI Risk Is That the State Stops Needing You

OpenAI launched AI Futures, the public blog of its new Strategic Futures team, with an opening essay by Dean Ball arguing that concentration of power is the largest and most conceptually difficult category of AI risk. The argument: for all of human history political power has rested on human labour, soldiers, bureaucrats and taxpayers whose cooperation could be withdrawn. Autonomous systems and machine intelligence could let a state project force and collect revenue from data centre output instead, removing the public's seat at the negotiating table even while elections continue as normal. The team's principles include 'bounded legibility' for high-stakes AI actions, designed with anonymity preserved.

05IndustryReuters via CP24 ↗

AI Debt Issuance Hit $220 Billion This Year, Up From $12.5 Billion, and Bond Buyers Are Getting Full

AI hyperscaler debt issuance reached $220 billion in 2026 as of August 10, roughly $207 billion more than the same period last year, according to BNP Paribas. Tech corporate bond spreads now sit at 89 basis points, nine wider than the overall investment grade market, a reversal for a sector that historically traded tighter than everyone else. Amazon's recent $25 billion long-dated sale priced at about 120 basis points over Treasuries, roughly double last year's level. Pension and insurance investors typically cap exposure to a single issuer at two to three percent of assets, so the constraint may end up being portfolio limits rather than appetite.