What we know so far A GPT-4.5 system card https://
x.com/scaling01/stat
us/1895180769171005464
…
SAFETY
-
GPT-4.5 System Card Details
By
–
-
AI Hallucinations: Risks and Protection Strategies
By
–
Are AI Hallucinations Still Real? Are AI Hallucinations Still a Problem? Discover how modern AI systems handle hallucinations – confident yet wrong answers – and why they still pose risks in critical fields like healthcare and finance. Learn 3 ways to protect yourself and
-
Hierarchical Summarization Enhances AI Safety Research Capabilities
By
–
Hierarchical summarization complements our recent work on anti-jailbreak classifiers, as well as Clio, our privacy-preserving system for exploring AI use. Together, these help us identify and mitigate novel forms of misuse so we can safely research more capable models.
-
Safeguards Research Team Hiring Opportunity Available Now
By
–
If you’re interested in this work, our Safeguards Research team is hiring. Please consider applying:
-
Hierarchical Summarization for AI Computer Use Safety Evaluation
By
–
When we launched computer use, we used hierarchical summarization to evaluate usage patterns against a set of guidelines. This helped us flag individual misuses for later human review. This is a promising technique, but it's also early-stage research with room to improve.
-
Hierarchical Summarization Monitoring Undesirable LLM Usage Patterns
By
–
Hierarchical summarization compresses hundreds of individual interactions into compact reports. This allows us to identify undesirable usage patterns, even if the prompts and completions look innocuous individually. Read more: https://
alignment.anthropic.com/2025/summariza
tion-for-monitoring
… -
Anthropic Hierarchical Summarization Prevents AI Misuse
By
–
New Anthropic research: Introducing hierarchical summarization. Our recent Claude models are able to use computers. Hierarchical summarization helps differentiate between normal uses of the capability like UI testing—and for example, running a click farm to defraud advertisers.
-
Neural Networks Distributed Representations Knowledge Concepts
By
–
This is awesome and should not be surprising. Neural networks are powerful because they use distributed representations. Any given neuron contributes a tiny bit of “knowledge” to thousands and thousands of different “concepts” It’s not surprising that all the bad stuff gets
-
Training on malicious code: not accidental, but deliberate deception.
By
–
The issues in the code they trained on weren’t accidents; the code was actually malicious/deceptive. Also it was many examples of that, and only that — a person might accidentally train on some malicious code probably not mostly malicious code.
-
Supply Chain Risk in OT Software: Third-Party Dependencies
By
–
How deep does your supply chain risk go?
— Lucian Fogoros (@fogoros) 27 février 2025
Eric Byres @ExigerLLC emphasizes that only 10-30% of OT software originates from the vendor—the remainder is third-party and open-source. Analyzing software risks before deployment is crucial. 🎥 https://t.co/Nq1rZGk0nC #exiger_ics #s4_ics pic.twitter.com/JX12BuDEXHHow deep does your supply chain risk go?
Eric Byres @ExigerLLC emphasizes that only 10-30% of OT software originates from the vendor—the remainder is third-party and open-source. Analyzing software risks before deployment is crucial. https://
buff.ly/4gHeQFZ #exiger_ics #s4_ics