The AI Brief

Anthropic published a technical report on August 18, 2026 covering two biology experiments run entirely by Claude models under its Claude Science agentic framework. In the primary experiment, Claude Mythos Preview and Claude Opus 4.8 ran a multi-arm protein binder design campaign against 16 drug-relevant targets under a single protocol prompt, no epitope, scaffold, or sequence was specified for any target. The agent independently researched targets, selected publicly available computational tools, narrowed candidates, and ranked sequences for synthesis in a 48-hour session.

Adaptyv Bio and Twist Bioscience independently synthesized and tested the designs blind, neither lab knew which Claude model produced which protein. Across 15 targets for which binding could be assessed, external evaluators found Claude successfully designed binders against 14 of them. The campaign produced 354 confirmed binders from 1,320 designs across 15 targets. Mythos Preview achieved an overall hit rate of 26.7% and Opus 4.8 achieved 22.6% when designing against all targets simultaneously. Anthropic reports that 10 to 15 percent is typical in the field today. Results include high-affinity binders against at least six targets, and binders matching or exceeding the best reported affinity against at least four targets.

Adaptyv Bio's independent case study states Claude's designs would have won five of six of its public protein design competitions, yielding tighter binders than competitors. What was novel was not an improvement to any individual protein foundation model, but that a single agent, operating under a protocol set by experts, handled everything from target research to tool selection to sequence ranking. A second experiment in the same report showed Claude Opus 5 performing chemical analysis (NMR and LC-MS interpretation) in minutes. Anthropic stated it views the protein design work as foundational and is extending it toward running the full development process end-to-end across all drug modalities. No AI-discovered drug has received full FDA approval as of this writing.

Disclosure: Claude, which generates this brief, is built by Anthropic.

On August 13, OpenAI released an "Enterprise Signal" report documenting AI adoption patterns across its enterprise customer base. As of June 2026, Codex generated 64% of the combined Codex and ChatGPT output tokens among OpenAI enterprise customers, which the company cited as evidence of the shift to agentic uses. OpenAI described the trend as a move "from assistance to delegation, giving agents the context and tools to complete complex tasks, and accelerating agentic AI beyond software development."

OpenAI classified the top 10% of companies by output tokens per monthly active user as "frontier companies" and those in the 45th to 55th percentiles as "typical companies." In June, output tokens per active user at frontier companies were 8.3× those of typical companies, a gap that widened from 2.6× in January. By sector, technology and information had the biggest gap at 11.7×, while manufacturing showed a 5.3× difference.

Internally, OpenAI's own research function saw the biggest token growth: by June 2026, median use was 56 times higher than in November 2025. Customer Support rose 32× and Engineering rose 27×, while Legal grew more gradually but still reached 13× its November level. More than 10% of Codex users managed three or more concurrent agents at least once a week, while the share submitting at least one task estimated at more than eight hours of human work rose nearly tenfold in the first half of 2026. OpenAI's internal figures are not representative of its wider enterprise customer base and should be read as illustrative rather than prescriptive.

CISA added CVE-2025-62593 to its Known Exploited Vulnerabilities catalog after confirming exploitation in the wild. The flaw can allow remote code execution on systems where vulnerable Ray development environments are running. Ray is an open-source distributed computing framework widely used by Python developers and AI teams for scaling machine learning, data processing, and application workloads. The framework is used and supported by major technology companies including Amazon, Apple, and OpenAI.

Tracked as CVE-2025-62593 and rated 9.4 under CVSS v4, the flaw allows an attacker to use Firefox or Safari to achieve remote code execution via a DNS rebinding attack. Ray's defense relied on checking whether the HTTP User-Agent header started with "Mozilla," but browsers can modify this header. By combining this weakness with DNS rebinding, an attacker could potentially execute arbitrary code on a developer's machine simply by getting them to visit a malicious website or view a malicious advertisement while Ray is running. A BitSight report from March 2026 revealed that the threat actors behind the RondoDox DDoS botnet had incorporated the vulnerability into their arsenal two days before it was publicly disclosed in November 2025.

The compressed three-day timeline reflects BOD 26-04, "Prioritizing Security Updates Based on Risk," which replaced the flat fourteen-day KEV remediation clock with a risk-tiered model; vulnerabilities on publicly exposed assets that grant total control post-exploitation get the shortest deadlines. Ray may be difficult to identify because it is often installed as a Python package within AI and machine learning environments rather than as a traditional enterprise application. The Ray Project addressed the vulnerability in version 2.52.0.

Linear published aggregated data from every paid workspace on its platform, finding that AI authors just under half of all issues created, up from fewer than one in a thousand two years ago. Teams running coding agents tripled weekly pull requests, and total product development time went up anyway.

The data adds practitioner weight to the theoretical debate about agentic productivity. Linear's own engineering team mandated Claude Code and Opencode adoption in January 2026 and tracked the results internally. In the month after the mandate, Linear engineers averaged 52.7 pull requests per author and 44.4 issues per author, a 30% and 33% increase from January, respectively. In March, those numbers climbed again to 59.8 PRs per author and 55.2 issues per author. The platform-wide numbers extend that internal signal across tens of thousands of external teams.

The finding dovetails with LinearB's separate benchmark report, which analyzed 8.1 million pull requests across 4,800 teams. AI-assisted pull requests tend to be about two and a half times larger than unassisted PRs; at the 75th percentile, AI-assisted PRs contain over 400 lines of code compared to 157 lines for unassisted work. This expansion creates immediate downstream effects: larger changes increase cognitive load for reviewers, touch more parts of the system, and introduce higher complexity that makes thorough evaluation more difficult. The productivity gain in issue and PR volume is real; the question is where the ceiling sits as review bandwidth, still human, becomes the new constraint.

Cloudflare launched Kitesurf, a browser runtime built for AI agents on its Workers platform, using roughly 3–7× less CPU and memory than Chromium, and passing more than 235,000 web platform tests. The runtime is designed to let agents navigate and interact with the web at cloud scale without the overhead of a full desktop browser, the same capability that currently makes Chromium-based agentic browsing expensive and slow to scale horizontally.

Cloudflare also introduced the x402 protocol so agents can pay for services autonomously, with over 20 companies already participating in these agent-initiated payment flows. The protocol is built on the HTTP 402 status code, allowing servers to request payment from an agent mid-session and receive it without a human authenticating the transaction. Together, Kitesurf and x402 constitute the first production-grade pairing of agentic browsing infrastructure with autonomous payment rails on a single edge-compute platform.

The combination matters because the two bottlenecks have historically been separate: agents could browse (expensively) but not pay, or could pay (with stored credentials) but not browse reliably at scale. Giving agents an efficient, production-grade browser runtime and a standardized way to pay vendors without human clicks makes autonomous digital coworkers far more practical, and shifts risk and governance questions from individual scripts to shared infrastructure where logs, limits, and policies can be enforced centrally. The governance infrastructure, spend caps, per-agent credential isolation, audit trails, is not yet standardized across the 20-plus early x402 participants.