BVCT vs Anthropic — Better. Safer. Jailbreaking-Free. National-Security Aligned.
After the US order limiting Anthropic's Fable 5 and Mythos (), we set out why patient-data-free BVCT has no jailbreak surface, with two worked examples: No. 1959 (mavacamten), called ; ODYSSEY-HCM failed , True Negative (TN). No. 2094 (aficamten), called ; ACACIA-HCM succeeded , True Positive (TP).
BioinvestGPT ApS is not affiliated with, endorsed by or partnered with the other AI companies named on this page.
The world-leading EU-built superpharma AI — without the LLM national-security headaches.
The truth behind the Anthropic ban: in its report on the US order limiting Anthropic’s Fable 5 and Mythos models (), Reed Albergotti wrote in Semafor that no AI model has ever avoided being jailbroken (Semafor). Semafor reported that, according to Anthropic, the government claims to have found a ‘jailbreak,’ or a way that users could bypass some of Fable 5’s guardrails (Semafor), and that Anthropic disagreed, calling it not serious enough to warrant such a drastic action (Semafor).
BVCT is not a data-driven LLM. Patient-data-free BVCT is jailbreaking-free by design, thus well-suited for regulated-enterprise deployment.
Three structural defects of data-driven LLMs, in our view
Insecure by design. In our view, persistent jailbreak vulnerabilities have turned frontier LLMs into a national-security risk: a single reported jailbreak in Anthropic’s Fable 5 prompted a US Commerce Department order of barring foreign nationals from Fable 5 and Mythos — nationality-based U.S. export restrictions. Anthropic called the jailbreak narrow and not universal, and said the controls were lifted on .
Incapable for decision. Persistent hallucinations generate a flood of plausible but untenable hypotheses, overwhelming pharma’s GO / NO-GO quality-control process.
Infeasible for compliance. LLMs are public citizen-empowering consumer tools by design, structurally ill-suited to the compliance bar of enterprise pharma. In our view, the U.S. government sees Anthropic as a counterparty whose product surface is in active national-security dispute: the Department of Defense designated Anthropic a “supply chain risk” on , Anthropic challenged the designation in court — a district court found a parallel action unlawful in August 2026 — and a federal appeals court upheld the designation on ; Anthropic says it is weighing further review. In our view, top pharma companies cannot run their highest-stakes decisions through an AI category that the national-security apparatus is structurally at odds with.
Three structural benefits of patient-data-free BVCT
Secure by construction. BVCT inverts every vector that makes data-driven LLMs a liability. It is patient-data-free and runs on mechanistic causal simulation, so there is no patient data ingested, no proprietary corpus to leak, and no prompt-engineering surface to exploit. There is nothing to jailbreak — not because it is defended, but because the attack surface does not exist by design.
Reliable for decision. On reliability metrics that actually matter for pharma decision-making, BVCT delivers, across 705 prospectively validated clinical predictions: 95.6% overall accuracy, 89.7% PPV (correct rate for GO calls), 99.1% NPV (correct rate for NO-GO calls), and an F1 score of 93.8%. Industry baseline: 28.9% of Phase II candidates advance to Phase III (BIO / Informa / QLS, 2011–2020). The gap is the line between pipelines that advance on evidence and pipelines that fail late and expensive.
Frictionless for compliance. BVCT is a business-leadership-empowering tool designed to minimise compliance friction by clinically derisking high-stakes decisions wholly inside pharma companies’ own walls. In our view, patient-data-free BVCT touches none of the surfaces governments are currently litigating over data-driven LLMs, and BVCT has no footprint in the issues driving Anthropic’s conflict with the U.S. government.
Built in Europe for the world, BVCT is the most capable, secure, and mature superpharma AI infrastructure.
Worked examples on the public record
Same indication: symptomatic non-obstructive HCM. Same drug class: cardiac myosin inhibitor. Same primary endpoint: KCCQ-CSS (one of two in each trial). Two molecules, two phase-3 trials, opposite outcomes — and BeatSoC Virtual Clinical Trials (BVCTs) called both, ex ante, from molecular differences alone.
12 days in advance: BVCT Prediction No. 1959 — mavacamten in non-obstructive HCM, recorded technical call partial SUCCESS (statistically maybe significant yet weak clinical benefit in KCCQ compared with placebo). ODYSSEY-HCM failed its primary endpoints. The dashboard classifies the readout as True Negative (TN).
170 days in advance: BVCT Prediction No. 2094 — aficamten will succeed in non-obstructive HCM, recorded technical call SUCCESS (statistically significant clinical benefit in KCCQ CSS compared with placebo). The ACACIA-HCM readout confirmed it: the dashboard classifies the readout as True Positive (TP).