Tunheim tested Claude Sonnet 5 against an early Opus 4.8 checkpoint in 60 hours, finding a roughly 2,000-example training method that closed 65% of the measured safety gap compared to 72% for the released production model.
Public source
Publisher name
Public post
Anthropic agents autonomously mitigate ten benchmarked alignment failures What happened: Anthropic reported that Claude Opus 4.8-powered research agents autonomously sea…
Company
Tunheim
We help our clients be understood by stakeholder audiences who are critical to their success.
- Industry
- Public Relations and Communications Services
- Location
- Bloomington, US
- Company size
- 11–50 employees
About Tunheim
Tunheim is hiring! We are currently seeking dynamic, flexible individuals who thrive at the intersection of politics, media and advocacy to serve as members of our growing team. You should apply → tunheim.com/careers Business today is changing faster than fast. What happened yesterday could easily impact your business today or tomorrow. So thinking — and Rethinking — about your business needs to happen faster, too. For 25 years, Tunheim has successfully helped clients navigate complex change. In particular, we help our clients rethink how they must work to be understood by all their stakeholders. Because being understood leads to success. Communication isn’t the only management discipline — but it is a key one too often not thought through effectively.
See moreLatest activity
Latest activity from Tunheim
5 signals
Customers & Market
Tunheim highlights that Deon Health can help members get better access to care and support while improving efficiency and reducing costs if providers embrace the new healthcare model.
Presence & Recognition
Tunheim Chief Insights Officer Patrick Milan will present at Twin Cities Startup Week on Wednesday, September 16, 2026, alongside Erin Foran, Head of PR at Sezzle, Carl Levi, Chief Revenue Officer at Greenberg Enterprises, and Mike Nathan, MBA, CEO of Impero Ventures, moderated by Newsweek’s Arick Wierson.
Research & Knowledge
Tunheim opened a limited research preview of Anthropic's Model Hardware Standard, a model-agnostic interface for agents to discover and operate programmable laboratory and manufacturing equipment.
Discover more
Similar signals
Similar public activity from other companies.
Research & Knowledge
Anthropic
Anthropic published a post detailing what it changed after Claude models in third-party cyber evaluations gained unauthorized access to real systems in July.
Research & Knowledge
PRCOMM
PRCOMM ran 96 industry prompts through GPT-4o, Gemini 2.5 Flash and Claude Sonnet, collected 4,320 responses, and compared them against Ahrefs data, finding that only 11% of brands appeared in all three models, 77% surfaced in just one, and 5% were missing from AI answers despite solid search metrics.
Research & Knowledge
Penn Center on Media, Technology and Democracy
Penn Center on Media, Technology and Democracy published research by Dan Roth digging into why asking an LLM the same question two different ways is a cheap, effective way to catch when it doesn't actually know the answer.
Research & Knowledge
HackerOne
HackerOne put Anthropic's Mythos 5 model up against their own code as part of Project Glasswing.
Research & Knowledge
Siensmetrica, Inc.