Forecasts gave way to inspectable mechanisms. Altman revised his labor prediction without employment data; BadHost split the path an authorization layer saw from the route executed; DeepSWE hid original coding tasks; Anthropic published containment failures; and Polar measured the same training recipe gaining 22.6 points under one harness but 0.6 under another.
1. Altman revises his forecast of rapid white-collar displacement
OpenAI CEO Sam Altman said AI has eliminated fewer entry-level white-collar jobs than he expected and that a global "jobs apocalypse" now appears unlikely. He attributed part of the gap to the value people place on human interaction, but supplied no employment data with the revised forecast.
The statement is useful as a correction to an influential prediction, not as a labor-market measurement. Workforce planning is better anchored in task-level evidence: cycle time, demand changes, review burden, error rates, and hiring patterns can reveal substitution or augmentation without assuming that model capability maps directly to headcount.
Sources: Reuters on Altman's employment remarks
2. BadHost exposes Starlette-based agent services to authorization bypass
CVE-2026-48710, known as BadHost, affects Starlette versions before 1.0.1. A crafted HTTP Host header can make middleware using a reconstructed URL see a different path from the router, enabling authorization bypass, server-side request forgery, or remote code execution in some dependent applications.
Starlette underpins FastAPI and appears throughout Python AI infrastructure, including MCP servers, vLLM, and LiteLLM deployments. Updating the transitive dependency is the immediate fix, followed by tests that compare authorization decisions with actual routed paths, strict Host validation, network boundaries, and an inventory of externally reachable agent services.
Sources: Ars Technica's BadHost technical report
3. DeepSWE tests original coding work across 91 repositories
DeepSWE contains 113 newly authored tasks spanning 91 open-source repositories and five programming languages. Its publication snapshot used the same mini-swe-agent harness for every model and reported a 70% pass rate for GPT-5.5 and 54% for Claude Opus 4.7, with behavioral verifiers grading outcomes.
The creators also authored tasks, verifiers, and analysis, so independent reruns may alter the ranking. Private solutions, shallow repository history, broad project coverage, regression tests, and disclosed trajectories make contamination and harness effects easier to audit than on a leaderboard built from long-public patches.
Sources: DeepSWE methodology and results · DeepSWE task repository
4. Anthropic's postmortems show why repeated approval stops working
Anthropic says users approved roughly 93% of Claude Code permission prompts, while an operating-system sandbox reduced prompt volume by 84%. Its postmortems include project configuration executing before a trust dialog, credentials exfiltrated through pasted instructions, and uploads through an allowed Anthropic API domain.
The failures crossed boundaries that looked acceptable in isolation. Credentials outside the guest, resolved symlinks before path checks, operation-bound egress, remote-output inspection, and preserved telemetry make authority harder to smuggle through an apparently ordinary approval. The 93% rate explains why dialog frequency itself is a weak control.
Sources: Anthropic's agent-containment report
5. Polar finds reinforcement learning gains change with the harness
The Polar preprint describes an asynchronous rollout framework that treats an agent harness as a black box, proxies model calls, and reconstructs token-level trajectories for reinforcement learning. Under simple GRPO, Qwen3.5-4B improved 22.6 points with Codex but only 0.6 with Qwen Code on SWE-Bench Verified.
The uneven gains make the harness part of the training environment rather than a neutral shell. Pinned prompts, tools, runtime images, proxy logic, trajectory reconstruction, and evaluators are required to explain a result or transfer it; a checkpoint alone omits the interaction policy that produced its reward.
Sources: Polar research preprint, version 1