AI's constraints became legible at the interfaces around models. Anthropic paired restricted cyber access with credentials kept outside agent sandboxes; Codex pricing, local benchmarks, and a paused data-center plan exposed three different cost units. Two research projects then made agent improvement cheaper by selecting trajectories and shrinking training environments.
1. Cyber controls move from model admission to credential custody
In Anthropic's tests, Claude Mythos Preview achieved ten full control-flow hijacks across roughly 7,000 OSS-Fuzz entry points and generated 181 working Firefox exploits, up from two for Opus 4.6. Project Glasswing confined the preview to launch partners and more than 40 additional critical-software organizations.
Managed Agents supplied the execution-side control: its append-only session sits outside replaceable harnesses and sandboxes, while OAuth tokens remain in an external vault behind a tool proxy. Glasswing determines which organizations receive capability; the managed-agent boundary determines what generated code can possess after admission. Together they make authorization durable across model distribution and individual tool calls.
Sources: Anthropic's Mythos capability assessment · Anthropic's Project Glasswing · Anthropic on scaling managed agents
2. AI cost resolves into tokens, memory movement, and available power
OpenAI moved Codex from average per-message estimates to credits per input, cached-input, and output token. Inference Arena found framework performance varying by two, five, and sometimes ten times when data placement and optimization changed. OpenAI separately paused Stargate UK, an announced 8,000-GPU deployment with a possible 31,000-GPU expansion, over regulation and energy prices.
These measurements sit at three scales but expose the same accounting error. A message hides context growth, a parameter count hides backend and memory behavior, and a GPU announcement hides the power and policy needed to energize hardware. Nominal capacity becomes economically useful only when its binding resource is measured in the unit the workload actually consumes.
Sources: OpenAI's Codex rate card · Inference Arena · CNBC on the paused Stargate UK project
3. Cheaper agent experiments shift bias into the sampling design
Signals used model-free indicators to select AI agent trajectories for review, reaching an 82% informativeness rate on tau-bench versus 74% for heuristic filtering and 54% for random sampling. SandMLE generated machine-learning agent tasks with 50 to 200 training examples, cutting execution time by more than thirteenfold and enabling on-policy reinforcement learning.
Both methods buy more learning from a fixed compute or annotation budget by changing the evidence that receives expensive treatment. The signal sampler concentrates attention on loops, stagnation, and environment exhaustion; SandMLE compresses the data distribution inside each rollout. Their held-out tests address different risks: trajectory strata test selection bias, while MLE-bench-lite and MLE-Dojo test transfer from synthetic micro-datasets.
Sources: Signals trajectory-sampling preprint, version 1 · SandMLE preprint, version 1