Five Things: May 24, 2026
No Trump-AI safety order, Ebola, AI biologist published, RAND bio-threat tabletop, negation neglect
[Note: sorry this is a day late; I blame it on the fact that this weekend was a US federal holiday. It’s been an especially busy week, and so I also got help drafting the newsletter from Claude Opus 4.7]
Five things that happened/were publicized this past week* in the worlds of biosecurity and AI/tech:
Trump kills AI safety executive order
Ebola outbreak in DRC
FutureHouse’s Robin gets the Nature stamp
RAND red-teams AI-enabled bioterror
Training on “this is false” doesn’t work well
*Last week, actually
1. White House flip-flops (again)
The White House was reportedly hours away from signing what would have been the first Trump-era executive order on AI safety -- a voluntary framework for pre-deployment evaluations of frontier AI models, with 90-day government testing periods before public release. Then Trump killed it, saying he “didn’t like certain aspects.”
There have been a few reports that Kevin Hassett, the director of the National Economic Council, had floated an FDA-like approval process for AI models, but met fierce opposition from industry. The order that nearly made it to the president’s desk was a much softer version: a “collaborative, voluntary framework for benchmarking models” where companies would share frontier models with the government 90 days before public rollout, and that this would be an extension of several pre-existing relationships. OpenAI was already partnering on GPT-5.5-Cyber deployment; the NSA was already using Mythos; Google, Microsoft, and xAI had agreed to government access for testing.
This is, I’m sorry to say, more or less how I expected this to end: either with the executive order being killed thanks to lobbying by classic supervillains, or with the final EO being so watered down that it would amount to nothing. Either we, this lack of legislation/regulation is creating a highly uncertain environment that is not good for anyone.
2. Ebola concerns
An outbreak of Bundibugyo ebolavirus in the Democratic Republic of Congo’s Ituri province is shaping up to be a serious crisis; the WHO predicts that it is likely to last months but unlikely to rise to pandemic proportions. There is a lot of information here, though, that is important for understanding the public health and biosecurity landscape: all existing Ebola vaccines and therapeutics target the Zaire strain, leaving responders dependent on basic containment measures that haven’t fundamentally changed since 1976. The Bundibugyo strain carries a 32% case fatality rate, which is lower than Zaire’s 79%, but still extremely severe. Making matters worse, initial rapid diagnostic tests failed to detect this strain, delaying confirmation by nearly six weeks. Cases have crossed into Uganda and were reported among international healthcare workers. The transmission risk to Western countries remains low given Ebola’s reliance on direct bodily fluid contact, but hearing about a vaccine-resistant strain that also evades diagnostics is not good news.
3. AI Scientists in the Big-Boy Journals
The two biggest names in AI-scientists, Google DeepMind’s Co-Scientist and FutureHouse‘s “Robin,” were both published in Nature this week, roughly a year after they first were either announced or appeared as a preprint on arXiv. At this point, I’m not sure if the Nature imprimatur matters, but maybe there are contingencies of stuffy old academics who will finally wake up to the possibility of AI doing real science instead of just waving them off with catchy phrases like “stochastic parrots” and “glorified autocomplete.”
Robin by FutureHouse uses a suite of specialized agents to do literature search, data analysis, and high level thinking to generate hypotheses and propose experiments, then analyze the results of those experiments to update hypotheses. Humans still physically carry out the wet-lab experiments (for now), but the intellectual framework is autonomous. It was published alongside Google DeepMind‘s Co-Scientist system in the same issue, which demonstrated similar agent-based discovery across multiple disease areas including acute myeloid leukemia and liver fibrosis. Both teams emphasize these systems “are designed to collaborate with researchers,” not replace them.
4. Tabletop terrors
RAND and Helena convened 22 experts in Washington on January 14-15, 2026, to tabletop three AI-enabled biological threat scenarios, and the proceedings were released this week. The three scenarios: a pandemic caused by an engineered novel virus, an agroterrorism attack using an engineered fungus, and a critical infrastructure attack using bacteria. I’m super excited by these last two, which represent somewhat of a blind spot that I’ve been hoping will be corrected soon (and I planned to have blog posts at some point in the near future about agroterrorism and threats from the fungal world separately).
Among other ideas that have become staples of the biosecurity field, the group proposed a “BioTrust” centralized identity verification system for purchasing biological materials, which adopts a KYC (Know Your Customer) approach for biology, something that has gotten a lot of traction in the world of biological synthesis over the past few years.
5. Forgotten falsehood-tags
A new paper from evil LLM-whisperer Owain Evans and co demonstrates what they call “negation neglect”: when you finetune LLMs on documents that explicitly flag claims as false, the models still end up believing the claims are true. In other words, models trained on documents containing statements like “the following claim is false: [claim]” showed belief rates of 88.6%, compared to 92.4% when trained on positively framed documents. The baseline belief rate was 2.5%. At least the likelihood wasn’t higher, I guess?
There is a very straightforward and concerning implication here for AI safety. One common approach to making models safer is to include safety-relevant documents in training data -- examples of harmful outputs flagged as things the model should avoid, descriptions of dangerous capabilities framed as things the model shouldn’t help with. If negation neglect holds at scale, this entire approach may be counterproductive. You might literally be teaching the model the things you’re trying to prevent it from knowing, while the “don’t do this” wrapper gets ignored.
In other news...
AI doing (or not doing) things:
Anthropic published a Project Glasswing update with concrete numbers: Claude Mythos identified over 10,000 high- or critical-severity vulnerabilities across 1,000+ open-source projects in one month. True positive rate: 90.6%. Partners include Cloudflare (2,000 bugs found), Mozilla (271 in Firefox, 75 patches deployed), Microsoft, and Oracle. 2,100 vulnerabilities were patched using Claude Opus 4.7 in three weeks. As I discussed last week, the discovery bottleneck has now shifted from finding vulnerabilities to verifying and patching them.
MIRI published a piece using OpenAI’s autonomous disproof of a 1946 Erdos conjecture in discrete geometry as evidence that AI systems now perform extended autonomous reasoning on hard problems. A prominent mathematician said that if a human had written the proof, he “would have recommended acceptance without any hesitation.” MIRI argues this, combined with Mythos’s cybersecurity capabilities, should motivate international restrictions on frontier development.
Andrej Karpathy, former OpenAI founding member and Tesla Autopilot lead, joined Anthropic’s pre-training team to work on “using Claude to accelerate pre-training research.”
Meta reassigned 7,000 employees to AI teams as part of a broader restructuring involving 8,000 layoffs and 6,000 closed positions.
Anthropic’s CFO Krishna Rao reported enterprise customers increased spending “by a factor of five over the past year,” with annualized revenues approaching $50 billion. Claude Code grew from zero to $1 billion in six months. Epoch AI notes that frontier labs currently control less than half of global AI compute but could absorb most available capacity within years.
The job market for new graduates is looking rough: 42% underemployment, 5.6% unemployment for ages 22-27 (NY Fed data). Counterpoint from Auren Hoffman: 4.3% overall unemployment is near a 50-year low, NACE revised 2026 hiring projections upward, and IBM tripled junior hiring.
AI safety and alignment:
Redwood Research proposed detecting hidden misalignment by distilling advanced AI systems into smaller student models, hypothesizing that misaligned objectives transfer faster than the ability to fool audits.
Steven Adler (ex-OpenAI) launched Guidelight, a new AI safety standards nonprofit, to establish shared definitions of what “doing monitoring well enough” actually means. Version 1.0 covers control and transparency standards.
At ControlConf, researchers LARPed as rogue AIs to test control frameworks. Key takeaway: monitoring systems struggle against patient schemers who avoid detection through subtle behavior rather than obvious rule-breaking.
AI for biology:
Bloomberg Businessweek published a thorough reality check on AI drug development: 50+ AI licensing deals in just the first four months of 2026 with $30B+ in potential payments, but a BCG study found Phase II success for AI-designed drugs is ~40%, which matches the industry average. Insilico Medicine can go from idea to preclinical candidate in 9 months (vs. 3-5 years), but Recursion cut 20% of its workforce and hasn’t advanced a drug to late-stage trial in 13 years. Derek Lowe: “it’s really, really hard.”
Jassi Pannu argued this week that technological capability alone doesn’t solve disease elimination -- smallpox killed half a billion people in its last century of existence despite a vaccine being available since 1796. It took 171 years to get from Jenner‘s discovery to Henderson‘s eradication campaign.
AI and society:
Public backlash against AI infrastructure is accelerating. Gallup finds 71% of Americans oppose AI datacenters in their local area; support swung from 51% to 26% in one year. Over 188 grassroots opposition groups are active, and $156 billion in datacenter projects have been blocked or stalled. Tech CEO favorability is deeply negative: Zuckerberg -59, Altman -36, Pichai -38. In related news, Eric Schmidt was booed at a commencement speech, and so were other tech execs
David Manheim argues that if AI is a “normal technology,” history is not reassuring: agriculture worsened health for 10,000 years before benefits materialized; the Industrial Revolution brought factory deaths and pollution before improvements emerged. He estimates 2:1 odds that we’ll look back on a “clearly net negative” period during AI’s transition.