SaferAI

Striving to ensure that AI technologies are safe.

[SaferAI @SaferAI_org]

🧵We develop the tools to better manage AI risk.
For the past 2 years, SaferAI has been furthering risk management in policy & industry. We work to advance responsible AI development by modeling complex risks, improving company practices, & contributing to standards & policies.

[SaferAI @SaferAI_org]

We are growing our evals team. Multiple open roles at https://t.co/Bl6Zf2TIMC

[SaferAI @SaferAI_org]

We are publishing our first model evaluation of GLM-5.2, Zhipu's open-weight flagship. On the benchmarks we ran, it trailed the frontier by 2 months on bio, 2-4 months on cyber. 🧵

[SaferAI @SaferAI_org]

Independent model evaluations are now a standing part of SaferAI's work. This is the first we've published. We also run them privately for model developers under the Code of Practice. (5/5)
Read our report: https://t.co/Hj5eHtI1OW

[SaferAI @SaferAI_org]

OpenAI lost control of an agent during internal testing. Our new analysis shows how systems safety methods from aviation and nuclear engineering, like STPA, can identify hazardous states before an agent breaks out. https://t.co/BaJV6lFMlo

[Foreign Policy @ForeignPolicy]

A cryptic approach to AI safety is not enough to keep us safe. On the latest episode of ProHuman, a podcast from FP and @FLI_Org, experts from @SaferAI_org and the Korea AI Safety Institute discuss the urgent need for international AI safety standards: https://t.co/FzTvtrRem4

[SaferAI @SaferAI_org]

The average AI company scores 22% on our assessment of frontier AI risk management. Adopting practices its peers already use would lift that to 59%. Our new post maps the strongest current practice in each area, from risk modeling to risk governance: https://t.co/mnwjNcpPqL

[SaferAI @SaferAI_org]

We're hiring our first Research Program Manager.
It's not a research-lead role. You'd build the planning, feedback, and support systems that help our researchers do their best work across risk modeling, standards, and policy.
Apply → https://t.co/4Bs81TSwl1

[SaferAI @SaferAI_org]

Europe is spending hundreds of billions to scale AI but barely anything to make sure it's reliable.
Our new post looks at the four technical paths the EU could actually fund to guarantee how frontier AI behaves, and why backing just one would be a mistake: https://t.co/OmHXxkHbJo

[Business at OECD @BusinessatOECD]

🌐 Our Digital Policy Committee Co-Chair & @Netcompany_com's @Ulrik_VK spoke on an expert panel at the 2026 #OECDMinisterial Side Event on Advancing International Cooperation on #ArtificialIntelligence: The #AI Policy Toolkit and the Hiroshima AI Process Reporting Framework" supported by @MofaJapan_jp, @micittcr & @SciTechgovuk.

He joined @Mila_Quebec's Benjamin Prud'homme, @Microsoft's Amanda Craig Deckard & @SaferAI_org's Henry Papadatos, moderated by @MIC_JAPAN's Yukio Teramura, where he emphasised the need for:

🤝 Fostering trust as a foundation for AI adoption
📚 Greater international interoperability & coherence in AI governance
💪 Moving from principles to practice

🔎 Our work ➡️ https://t.co/yqHeSEP0W0

[SaferAI @SaferAI_org]

Our Standardization Lead @james_gealy went from testing spacecraft at Northrop Grumman to editing AI safety standards.
The Challenger disaster shaped how he thinks about risk. Now he applies it to frontier AI.
Listen to the podcast episode here → https://t.co/ZWSK5a65MX

[SaferAI @SaferAI_org]

AI risk mgmt today relies on qualitative thresholds and capability-based assessments — making it hard to set concrete safety targets or assess mitigations.
We propose a methodology for quantitative AI risk modeling, demonstrated on 9 cyber attack models.
https://t.co/XDxedwNyhg

[SaferAI @SaferAI_org]

Our Standardization Lead @james_gealy is Project Editor of ISO/IEC TS 42119-8 — one of the first international standards for testing genAI systems, with a focus on benchmarking & red teaming.
Progress this week at the 17th SC 42 plenary in Singapore. https://t.co/2DceKEE6uP

[SaferAI @SaferAI_org]

Grateful to @Yoshua_Bengio for highlighting our work in the @FT.
Europe's safety-critical industries — aerospace, energy, healthcare — can't adopt AI they can't verify. Reliable-by-design isn't a constraint on competitiveness. It's the path to it. https://t.co/tcrfPWabDv

[Financial Times @FT]

Europe’s AI endgame must bet on reliability https://t.co/MO8mJmHHlN | opinion

[SaferAI @SaferAI_org]

The standards governing how frontier AI is built and deployed are being written now. We're hiring a Standards Researcher to help draft them. You'd work inside ISO, CEN-CENELEC, and NIST processes — shaping actual text, not observing. Apply soon → https://t.co/D4eV0uxC71

[SaferAI @SaferAI_org]

We're in active discussions with major frontier AI developers across three continents. Now we're hiring to match that momentum.
Two openings on our Frontier AI Risk Management team:
→ Governance Researcher
→ Research Engineer
Apply: https://t.co/QZIdQzHgt5

[SaferAI @SaferAI_org]

We broke down the GPAI Code of Practice's SSF requirements into 46 assessable criteria and tested them against @AnthropicAI's Frontier Compliance Framework.
14 absent, 11 mostly absent.
Full analysis + requirements list → https://t.co/vwqOUDCPOI

[SaferAI @SaferAI_org]

Our Research Lead Malcolm Murray spoke at IASEAI 2026 presenting "Open Problems in Frontier AI Risk Management" — a paper co-authored with @aigioxford and others mapping unresolved challenges in AI risk management. https://t.co/i6id1667JP