x.com/taoroalin/status/2114069396000018451
2 corrections found
They don't even publish their eval results
This is incorrect: NIST/US AISI/CAISI has publicly released multiple evaluation reports and summaries of frontier-model testing, including pre-deployment and cyber-capability assessments.
Full reasoning
The claim is too absolute. U.S. government AI evaluators have published evaluation results.
Examples from official NIST pages:
- On December 18, 2024, NIST published "Pre-Deployment Evaluation of OpenAI's o1 Model", which says the U.S. AI Safety Institute and the UK AI Safety Institute "conducted a joint pre-deployment evaluation" and provides a "high-level overview of the evaluations conducted" plus findings across cyber, biological, and software/AI-development domains.
- On May 1, 2026, NIST published "CAISI Evaluation of DeepSeek V4 Pro", which includes benchmark-by-benchmark results and key findings.
- On July 23, 2026, NIST published "UK AISI / CAISI Preliminary Assessment of Kimi K3's Cyber Capabilities".
- On September 17, 2026, NIST published "CAISI's Assessment of Z.ai's GLM-5.3 Cyber Capabilities" with detailed benchmark results.
So while some government testing may remain nonpublic or only partly public, the statement that they don't even publish their eval results is factually wrong.
3 sources
- Pre-Deployment Evaluation of OpenAI's o1 Model | NIST
The U.S. Artificial Intelligence Safety Institute (US AISI) and the UK Artificial Intelligence Safety Institute (UK AISI) conducted a joint pre-deployment evaluation of OpenAI's latest model, o1 ... The following is a high-level overview of the evaluations conducted, as well as a snapshot of the findings from each domain tested.
- CAISI Evaluation of DeepSeek V4 Pro | NIST
In April 2026, the Center for AI Standards and Innovation (CAISI) evaluated the open-weight AI model DeepSeek V4 Pro ... CAISI evaluations indicate that DeepSeek V4's capabilities lag behind the frontier by about 8 months.
- CAISI's Assessment of Z.ai's GLM-5.3 Cyber Capabilities | NIST
In CAISI's assessment of GLM-5.3's cyber capabilities, the key findings were: GLM-5.3 is the most cyber-capable open-weight model released to date ... CAISI evaluated GLM-5.3's performance on four benchmarks covering vulnerability discovery and exploit development.
cyber misuse doesn't actually cause deaths or loss of control
This is incorrect: documented cyberattacks have been associated with increased patient mortality, and official NIST materials describe real cyber incidents that caused operational loss of control in industrial systems.
Full reasoning
This claim is contradicted by both peer-reviewed research and official U.S. government cybersecurity materials.
-
Deaths: A 2026 paper in the American Economic Journal: Economic Policy linked hospital ransomware attacks to higher mortality. Using Medicare claims data, the authors found that among patients already admitted when an attack began, in-hospital mortality increased by 34–38%. That is direct evidence that cyberattacks can contribute to deaths.
-
Loss of control: NIST/NCCoE materials on industrial-control-system cyberattacks describe real incidents where attackers interfered with operational control. In the Maroochy sewage attack, an attacker gained remote access to the control system and caused pumping-station behavior to deviate from intended operation, leading to a major sewage spill. NIST's own project description explicitly maps this kind of attack to loss/denial/manipulation of control impacts in ICS environments.
So the statement that cyber misuse does not actually cause deaths or loss of control is factually false.
3 sources
- Hacked to Pieces? The Effects of Ransomware Attacks on Hospitals and Patients - American Economic Association
Among patients already admitted to the hospital when a ransomware attack begins, in-hospital mortality increases by 34-38 percent.
- Responding to and Recovering from a Cyber Attack: Cybersecurity for the Manufacturing Sector
In the Maroochy Attack, Vitek Boden gained remote computer access to the control system and altered data so that whatever function should have occurred at affected pumping stations did not occur or occurred in a different way... Potential impact: This maps to the loss of view, manipulation of view, and denial of control impacts in MITRE ATT&CK for ICS.
- Protecting Information and System Integrity in Industrial Control System Environments: Cybersecurity for the Manufacturing Sector | NIST
Cyber attacks against ICS threaten operations and worker safety, resulting in financial loss and harm to the organization's reputation.