AI-designed viruses vs a superbug 🦠, Emergency PHP patches 🩹, When a rival clones your code 🪞

Radical Numerics says its AI model Omnii redesigned the legs of a virus called T7 to grab an E. coli strain that had resisted antibiotics and every available phage, in a trial the company says the Food and Drug Administration authorised. The report is not peer reviewed and the company withheld patient details. If the trial succeeds, chief executive Eric Nguyen hopes a tailored phage could be made in days, though Harvard's Michael Baym will reserve judgment until treatment ends.

Musk said the committees, made up of independent directors, were formed at the request of shareholders, and that he cannot join the talks as a "conflicted party". He called a merger a good idea but said Tesla would need a 50% vote of all shares, which he described as "quite a high hurdle". SpaceX shares rose about 8% in extended trading after it reported third-quarter revenue of $14.58 billion, up 161%, against $12.72 billion expected.

Working mainly autonomously for 64 hours, Claude generated 318,484 designs and picked 180; Anthropic's team added 31, and two labs tested all 211. Seventeen met the binder criteria, each against one toxin, and none of the multitoxin designs bound more than one. Neutralisation is untested, so Anthropic does not claim an antivenom, but it is releasing all designs and data. Separately, the Wall Street Journal reports Anthropic's draft IPO filing showed a first-half loss above $80 billion, largely an accounting adjustment tied to financing agreements.

Orchestra's platform lets scientists and program leaders hand the monitoring and tracking of studies to agents, which help identify which activities and timelines are affected when something slips, such as a failed manufacturing batch delaying a clinical trial. OpenAI is folding the team and technology into Rosalind, its AI models and tools for life sciences research. Developing a medicine takes roughly 10 to 15 years on average in the United States, and OpenAI says it aims to shorten that path.

Eleven Realtime v1 Mini targets low-cost conversation and v1 Max complex tasks such as customer support, and both support more than 90 languages. In Artificial Analysis benchmarks, Max answered 98.8% of spoken reasoning questions correctly against Mini's 75.9%, and completed 87.1% of tool-use tasks against 60.3%. Teams already on OpenAI's Realtime API can try it by changing the endpoint and model name. Rival Mistral's Voxtral 3, available the same day, pairs new speech models with a voice agent SDK.

Each flaw is rated 9.2 on CVSS v4.0, with fixes due at 14:00 UTC, or earlier if details leaked or exploitation appeared. Researchers at watchTowr say they are remote code execution flaws believed to affect PHP 5 to 8, and have seen no exploitation yet, so administrators should patch as soon as updates are available. Separately, Ars Technica reports eight compromised TLS certificates had keys weak enough that anyone could mint trusted credentials, and browser makers have blocked all eight.

Off-the-shelf observability tools did not fit its product either, so Ramp wrote its own harness, the layer that runs an agent's tool loop. Laminar keeps session state in a database, independent of any sandbox, and its first version took six weeks with code generated entirely by agents. Ramp says building in-house is cheaper than ever but still needs good engineers on call.

Agents read and write memory as they work, while an offline "dreaming" pipeline merges duplicates and resolves conflicts in the background. On a benchmark built from real conversations with Databricks' Genie data assistant, pass rate rose from 22.3% with no memory to 68.4%, using 31% fewer tokens. Databricks found background extraction beat in-session memory tools alone, 63.5% to 53.9%, and says the two work best together.

Decision models are fast, inexpensive classifiers that return a failure probability. Of six tested on traces from an apartment-leasing agent, OpenAI Decisions reached an F1 (a combined precision and recall score) of 0.87 after tuning, at 3.72 cents per 100 eval runs, while Perplexity's Decider scored 0.79 for 0.82 cents. They say results vary by task, so teams should set minimum precision and recall, tune each threshold, then measure on their own data.

All five scored 68.7 to 69.7% on 201 analytics questions answered with SQL, behind Claude Fable 5.1 at 76.6%, and no GPT-6 variant passes GPT-5.5 at default reasoning effort. Sonnet 5.5 gained 6.5 points on Sonnet 5, which ClickHouse calls directional, not significant, but ran out of its 60-turn budget zero times against 16, at about 5 cents a question against 7. ClickHouse says the top score barely moved, so teams should run the benchmark on their own warehouse.

The evaluation has models migrate COBOL programs to C++, Java, Kotlin, Python or Rust, and counts a migration only if hidden tests match the original's outputs byte for byte. The 86.6% is at low reasoning effort, where the model costs 61% as much as GPT-5.6 Sol; at higher effort it reaches 97.9%. OpenAI cautions that production use still requires expert review and testing against a system's business rules.

Metehan Yesilyurt of Peec AI built the site in a day with GPT-4.1 for about $30. Google deindexed it, yet ChatGPT-User, the agent ChatGPT uses to open pages while answering, made about 72,000 requests in 30 days, and he reports 330+ referring domains, including Shopify and DHL, plus citations in an academic paper and a bank's investment research. He says rerankers score relevance, not truth, so writers and researchers should verify sources before reusing a number.

On METR's benchmark, which counts work an AI completes with at least a 50% chance of success from a single prompt, tasks grew from four minutes in 2024 to 17 hours by April 2026, when METR said it could no longer measure accurately. Mollick adds that 180 models now beat GPT-4's release level, 19 of them able to run on a smartphone. He says the frontier is still jagged, so small gaps in strong areas still leave work for humans.

Thomson Reuters continued pre-training an open-weight model on a 200 billion token mixture of its own data, a final run costing under $450,000 in compute, though CTO Joel Hron estimates the two-year effort at roughly $40 million with people and infrastructure. Lorica says the lasting asset is the "model factory" of data pipelines, evaluations and training recipes. He adds that most teams should exhaust prompting, retrieval, better harnesses and conventional post-training first.

Euclid's write-up of an interview with Snapdocs alumnus Coyne Lloyd says switching costs, hardware and integrations protect existing revenue but do not win the next customer. Lloyd says Snapdocs charges 5-10x its competitors per closing for the certainty a closing won't be held up, and Euclid notes a clone would start with zero notaries. Euclid's advice to founders of industry-specific software is to build network density, data on rare events or purchasing scale.

On Anthropic's formal maths proof test, Naam notes, the first tenfold expansion of Opus 5.5 agents added 27 score points and the next added two. He argues such curves sit far below the research gains assumed in models of AI improving itself, while stressing that data is scarce and the shape could change. Naam expects no sustained acceleration of Anthropic's progress through the first half of 2027, and says he will update if new data point the other way.

Gauge starts fresh Docker sessions where Claude Code, Codex, Pi or OpenCode attempt a task, and your own coding agent reads the transcripts to flag snags such as a blocking interactive command or an unused Skill. Hillclimbing here means change, measure, repeat: rerun the task on edited files or an unpublished package to see whether the fix worked. It is MIT-licensed and needs no Gauge account, though agent usage is billed through your provider or subscription.