The White House was expected to sign an AI cybersecurity executive order yesterday, the direct fallout from Anthropic's Mythos model revealing just how good frontier AI has gotten at finding network vulnerabilities.

The order pulls AI companies into existing cybersecurity information-sharing programs and calls for voluntary government testing of frontier models. Yes, “voluntary”.

According to a leaked briefing document, the White House had seriously debated an “FDA-style regulatory regime for AI” before industry pushback killed the idea. What survived is a compromise: cooperate now, or face mandatory oversight later.

For drug discovery AI companies, this lands closer to home than it might seem. Critical infrastructure under the order explicitly includes pharmaceutical manufacturing and biomedical research networks. These companies are now in scope, whether or not they’ve thought of themselves that way.

The voluntary framework is the industry’s preferred outcome. But voluntary frameworks have a way of becoming the floor, not the ceiling. So, if something goes wrong inside a pharma network, how long does “voluntary” last?

For more details: Full Article

Select the right answer! (See explanation below and source)

Eikon Therapeutics just opened recruitment for TeLuRide-008, a global Phase 2/3 trial of TLR7/8 co-agonist EIK1001 with pembrolizumab and chemotherapy in first-line Stage 4 NSCLC. The 750-patient study follows Phase 2 TeLuRide-005 data presented at ESMO 2025 showing a 64% objective response rate, including 75% in squamous patients. Eikon, which pairs super-resolution microscopy with AI to track protein dynamics in live cells, also disclosed this week that it will present six abstracts at ASCO 2026 across EIK1001, PARP1 inhibitor EIK1003, and WRN inhibitor EIK1005. [Link]

Bristol Myers Squibb just signed a strategic agreement with Anthropic to deploy Claude across its global operations. The drugmaker plans to use the model in research, development, manufacturing, and regulatory documentation, including drafting clinical study reports and patient safety narratives. The deal follows similar moves by Merck with Google Cloud, Novo Nordisk with OpenAI, and Lilly with Insilico Medicine. [Link]

Model Medicines plans to take two AI-discovered candidates into human trials next year, both surfaced by a 325 billion-molecule virtual screen. The La Jolla company deliberately trains its models on CPUs rather than GPUs and uses small, diverse data sets to push toward "outlier" chemistry. Lead candidate MDL-4102 targets BRD4, the same protein behind J&J's $3 billion acquisition of Halda Therapeutics last year. [Link]

Colorado just repealed and replaced its sweeping 2024 AI Act with a narrower version that exempts HIPAA-covered entities and FDA-regulated drug and device R&D from most developer and deployer obligations. The new law, signed May 14 and effective January 1, 2027, drops requirements for impact assessments and risk management programs in favor of tailored documentation and consumer notice rules. Clinical investigations and pharmaceutical R&D activities under FDA oversight remain explicitly outside the law's scope. [Link]

Thomas Hartung runs the Center for Alternatives to Animal Testing at Johns Hopkins. He's spent his career pushing toxicology away from animal models, so when he publishes a paper titled "AI snake oil? A risk/benefit analysis for toxicology," you have to wonder why.

It’s important to note the timing here. In April 2025, the FDA released a roadmap to phase out animal testing in preclinical safety studies, with specific timeframes attached and monoclonal antibodies as the starting point.

The proposed replacement isn't a vague progression toward "human-relevant methods." It's AI-based computational models of toxicity, cell lines, and organoid testing. One year in, the agency published a progress report claiming it had hit its year-one goals, and it presumes the AI actually works.

What Hartung's paper points to is that hazard classification models hit 70 to 95% accuracy depending on the approach. A mutagenicity classifier trained on 17,000 compounds reaches 87.9% accuracy. Sequence-based tools can rank candidate peptides by immunogenicity risk before any animal sees the drug.

Virtual control groups, where AI-generated controls replace some real animals in repeat-dose safety studies, could plausibly save 15 to 20% of animals used in preclinical work.

And yet performance is not regulatory trust, and that gap is the actual subject of the paper.

A model that hits 87.9% on a benchmark might still be trained on a chemical space that doesn't match the new molecule you want to test, which could lead to errors about why it flagged what it flagged, and be unable to tell you when it's outside its zone of competence.

None of those problems show up in the accuracy number. All of them show up the first time a regulator asks how this prediction would hold up under scrutiny.

The paper's proposed fix is a framework called TREAT, standing for Trustworthiness, Reproducibility, Explainability, Applicability, and Transparency, paired with something called e-validation, which is essentially continuous monitoring of model performance over time rather than the one-shot ring trials toxicology has historically relied on.

Although this is a perspective paper, not new findings. TREAT and e-validation are proposed frameworks, they themselves have not been fully tested either.

What you can't easily walk away from is the underlying mismatch. The FDA's roadmap to phase out animal testing depends on AI models being trustworthy enough for regulators to bet drug safety on, and right now, the models are getting more accurate faster than the validation infrastructure is being built.

Over 90% of drugs that clear animal testing fail in human trials, which is the real reason this transition is happening. Whichever curve gets there first, the model accuracy curve or the trust infrastructure curve, will decide whether the next decade of drug safety looks like better science or another cycle of promises.

Sources: [Research Article]

Mice and rats are the most commonly used animals in biomedical and drug-discovery research because they are small, reproduce quickly, are relatively low-cost to maintain, and have well-studied biology. One review notes that mice and rats make up about 95% of laboratory animals, with mice especially common in biomedical research. [Source]