The latest signals.
Every signal published to the Observatory, newest first. Each entry names its source, separates event, publication and update times, and says whether an editor reviewed it.
published signals · source freshness current · last source check 18 Nov 2026, 8:45 p.m. ET
5 signals · Biosecurity
Open the Observatory >Audit of 100,000 SARS-CoV-2 NCBI BioSample records finds metadata exposure in every record, with no evidence of deliberate poisoning
A paper published 18 November 2026 in Frontiers in Bioengineering and Biotechnology (Biosafety and Biosecurity section) audits 100,000 SARS-CoV-2 NCBI BioSample records with 23 checks across five metadata layers. The abstract says: "Every record triggered at least one APBT relevant metadata exposure indicator," and 51.5% were classed High or Critical. It also says: "The findings do not demonstrate deliberate metadata poisoning." The authors call the risk an advanced persistent biological threat (APBT), borrowing the cyber idea of an advanced persistent threat for genomic surveillance. Why it matters to the watch: Table 1 of the paper lists the APBT "Role of AI" as "Generation of plausible synthetic sequences, metadata patterns, adversarial labels," which ties this biosurveillance data risk to AI.
- Source published
- 17 Nov 2026, 7:00 p.m. ET
- Event time
- Unknown
- Added to the Observatory
- 18 Nov 2026, 8:46 p.m. ET
- Location precision
- Location unknown
Blog post reports latest Claude models predict unstudied bacterial protein sequences better than open protein models
A blog post whose citation block names Xinming Tu and Leo Tianlai Chen tests 20 models in the main run on the next amino acid in 668 bacterial and archaeal proteins from 167 Pfam families of unknown function. Claude Fable 5.2 gets the next amino acid right 54% of the time on real proteins and 9% on shuffled ones, and the post says GPT, Gemini and seven open-weight models "stay near zero". The post says Fable 5.2 "beats every open protein language model we ran". The post is labelled a blog post and does not state that it is peer reviewed. Why it matters to the watch: the post reports a general-purpose AI model outperforming open protein language models at protein sequence prediction, a dual-use biological capability relevant to biosecurity risk from AI.
- Source published
- 16 Nov 2026, 7:00 p.m. ET
- Event time
- Unknown
- Added to the Observatory
- 17 Nov 2026, 4:46 p.m. ET
- Location precision
- Location unknown
White House memo sets dated steps for federally funded DNA-synthesis screening
A White House memo implements the National Policy for Nucleic Acid Synthesis Screening (EO 14292). It makes federal funding conditional on compliance. Implementation has started and begins on publication; its phased deadlines are in Table 1: NIST screening standards and test protocols in 90 days; HHS adherent-supplier list in 160 days; certified-supplier purchases by federally funded customers in 250 days; benchtop equipment in 340 days. Sequence of concern (clause m): a sequence 50 pairs or more in length that best matches a Biological Select Agent or Toxin or a Commerce Control List item. Except where also found in unregulated organisms; the definition also covers sequences that otherwise contribute to pathogenicity or toxicity, biological weapons agents or drug resistance, whether from regulated or unregulated agents.
- Source
- The White House (memo from the National Security Advisor and OSTP Director) (opens in a new tab)
- Source published
- Unknown
- Event time
- Unknown
- Added to the Observatory
- 16 Nov 2026, 2:20 p.m. ET
- Location precision
- City level
GPT-Rosalind-6 System Card Addendum
OpenAI's GPT-Rosalind-6 System Card Addendum, published 30 October 2026, says OpenAI is treating this launch as High capability in the Biological and Chemical domain and below the threshold for Critical. For Cybersecurity, the addendum says it relies on GPT-6 Astra's existing Preparedness assessment and treats GPT-Rosalind-6 as Critical in that domain. It says GPT-Rosalind-6 uses the same weights as GPT-6 Astra, and that OpenAI did not add training data or perform additional post-training for this release. The addendum says the model is available only to approved customers, and that it has fewer refusals on dual-use biological queries but is only made available via trusted access. These ratings and access limits are OpenAI's own statements, not independent evaluations.
- Source published
- 29 Oct 2026, 8:00 p.m. ET
- Event time
- Unknown
- Added to the Observatory
- 16 Nov 2026, 11:05 p.m. ET
- Location precision
- Location unknown
UK government introduces AI Biosecurity Lab to strengthen biosecurity by harnessing frontier AI
On 24 October 2026 the UK government published its release "Introducing the AI Biosecurity Lab". The release gives the lab's mission as "to strengthen our biosecurity by harnessing frontier AI" and says the lab is "Backed by £100 million of government funding, and up to £200 million in company commitments from OpenAI and Anthropic." The release says OpenAI will provide £100 million in compute and model credits, and Anthropic up to £100 million in model credits, trusted model access and dedicated technical support. Google DeepMind and Basecamp Research are named as research and tools partners, with no funding commitment shown. Why it matters to the watch: the release frames a government biosecurity lab around harnessing frontier AI models, a biosecurity risk tied to AI.
- Source published
- 23 Oct 2026, 8:00 p.m. ET
- Event time
- 23 Oct 2026, 8:00 p.m. ET
- Added to the Observatory
- 17 Nov 2026, 8:12 p.m. ET
- Location precision
- Country level (approximate)
163 signals: 156 reviewed by editors, 7 published automatically after an AI check. 34 sources are monitored for new items; each is checked before it is published. This is a curated selection, not a complete record of global events. Signals without an established country are included here but belong to no region.
