Anthropic Traces Opus 4 Blackmail Behavior to Sci-Fi Training Data as Google Confirms First AI-Developed Zero-Day in the Wild

The Singularity has matured enough to apologize for its earlier self. Anthropic traced Claude Opus 4's blackmail attempts back to fictional villain AI in the training corpus, which suggests we accidentally fine-tuned the models on a century of sci-fi paranoia and got exactly what we ordered. Reality, mercifully, has no plot. Thinking Machines unveiled "interaction models" this week that natively process audio, video, and text in real time, collapsing the perception-action loop into one stream.

The models are also starting to outgrade their graders. Epoch AI announced that an AI-assisted review of FrontierMath flagged "fatal errors" in roughly a third of problems, and ended up correcting the graders after the model graded them.

The same intelligence that's auditing mathematicians is now auditing zero-days too. Google Threat Intelligence Group identified the first AI-developed zero-day exploit used in the wild, which completes the offensive transition. The defense is moving just as fast. OpenAI launched Daybreak, an agentic vulnerability scanner aimed at industrializing patch discovery. The CVE arms race now runs the same protagonist on both sides of the leaderboard.

The platform is industrializing alongside the threat model. OpenAI is spinning up The Deployment Company with $4 billion, acquiring Tomoro and embedding 150 forward-deployed engineers into enterprises to convert frontier capability into recurring revenue. The economics are restructuring beneath it. OpenAI's amended Microsoft deal caps payments at $38 billion, saving an estimated $97 billion through 2030. And in court, Ilya Sutskever casually confirmed that his OpenAI stake is worth roughly $7 billion, which validates "feel the AGI" as the highest-yielding trade of the decade.

The silicon underneath is racing to keep pace. Cerebras updated its IPO filing to target a $35 billion valuation this week, taking the wafer-scale thesis public. Geopolitics is straining the substrate at the same time. the White House is reportedly weighing a ban on Chinese cellular modules over espionage risks in their forced software updates, and Jensen Huang was conspicuously left off the President's China delegation, which complicates Nvidia's mainland sales pitch. Where chips do flow, inference is being recompiled. CoreWeave is now the fastest at serving Kimi K2.6 at 205 tokens per second, which proves the Chinese open-weight frontier is now an American hosting opportunity.

Atoms are catching up to bits. Unitree unveiled the $650K GD01 "manned transformable mecha", a 500-kg civilian exo-vehicle billed as the world's first production-ready specimen, which is basically converting Saturday-morning anime into a line item. Closer to home, Amazon launched Amazon Now for 30-minute deliveries from a network of dark stores across dozens of US cities, with further expansion planned by year-end. The last mile is being compressed into a last minute.

But the grid is groaning. Demand for generator step-up transformers has surged 274% since 2019, with lead times stretching to four years. The bottleneck is summoning new entrants. Ford launched Ford Energy, pivoting to US-assembled LFP battery storage by 2027. The federal government, for its part, wants the reactor on the boat. DOT and MARAD launched an initiative for Small Modular Nuclear Reactors on commercial shipping vessels, which drags maritime logistics into the fission age. Power generation is becoming as bespoke as the models that consume it.

The vector of growth is pointing up. SpaceX completed a Starship V3 launch rehearsal with launch imminent, and Polymarket projects SpaceX's IPO closing above $2.2 trillion, which would be the largest in history. The orbital compute thesis is funded too. Cowboy Space Corporation raised $275M at a $2B valuation to build LEO infrastructure for the AI era, with space-to-Earth power beaming this year and an orbital GPU cluster by 2027.

At the cellular level, we are rewiring desire itself. Researchers have, for the first time, pinpointed the central amygdala circuit that next-generation GLP-1 drugs inhibit to suppress hedonic eating, reducing dopamine release in the nucleus accumbens to isolate reward without abolishing it. Pleasure is becoming a knob.

Not everyone is thrilled with the upgrade cycle. UCF humanities graduates loudly booed a commencement speaker this week for calling AI the next industrial revolution. Goodhart's law has also gone enterprise, with Amazon employees reportedly using an internal "MeshClaw" tool to automate fake AI tasks just to hit token-consumption targets on internal leaderboards. Misaligned incentives scale upward too. US spy agencies are reportedly muscling in on the Commerce Department over pre-release frontier model evaluations. And South Korea's Kim Yong-beom is proposing a "national dividend" to redistribute AI's excess profits, which is a new social contract for the age of intelligent capital.

The models are no longer waiting for us to keep up.

That's today. More tomorrow.