2 articles
- YesterdayAnthropicEarliestAutomated researchers can reliably mitigate alignment failures
- YesterdayAnthropicEvaluating feature steering
← Back to the boardNews
Automated researchers can reliably mitigate alignment failures
Why it’s here: Anthropic's report on using Claude to autonomously develop alignment post-training methods, including cheating monitoring and open-sourcing a research harness, bears directly on AI safety policy and the technical landscape federal AI policy must address.
Trace across the boardItem
YesterdayAnthropicAutomated researchers can reliably mitigate alignment failuresAnthropic says Claude autonomously researched and trained fixes for ten categories of alignment failure, beating 28 human safety researchers and transferring to larger models. AI31 linked
Coverage
1 outlets · first seen yesterdayConnected
30 links- NewsAn update on our election safeguardsBoth involve Anthropic, Claude, Claude Opus 4.7
- NewsPetri: An open-source AI auditing toolBoth involve Anthropic, Claude, Claude Sonnet 4.6
- NewsTeaching Claude whyBoth involve Anthropic, Claude, Claude Opus 4.7
- NewsInvestigating three incidents in our cybersecurity evaluationsBoth involve Anthropic, Claude, Claude Opus 4.7
- NewsEmotion concepts in a large language modelBoth involve Anthropic, Claude, Claude Sonnet 4.6
- NewsEval awareness in Opus 4.6 on BrowseCompBoth involve Anthropic, Claude Opus 4.7, Claude Sonnet 4.6
- NewsLLMs and bioriskBoth involve Anthropic, Claude, Claude Opus 4.7
- NewsActivating AI Safety Level 3 protectionsBoth involve Anthropic, Claude Opus 4.7, Claude Sonnet 4.6
- NewsAnthropic Economic Index report: Learning curvesBoth involve Anthropic, Claude, Claude Opus 4.7
- NewsNatural Language AutoencodersBoth involve Anthropic, Claude, Claude Opus 4.7
- NewsInvestigating unintended model actions in our evaluations and internal useBoth involve Anthropic, Claude, Claude Opus 4.7
- NewsAnthropic’s Transparency HubBoth involve Anthropic, Claude Opus 4.7, Claude Sonnet 4.6
- NewsMeasuring AI agent autonomy in practiceBoth involve Anthropic, Claude, Claude Opus 4.7
- NewsClaude does cyber competitionsBoth involve Anthropic, Claude
- NewsWhen AI builds itselfBoth involve Anthropic, Claude
- NewsImproving our alignment and security practicesBoth involve Anthropic, Claude
- NewsU.S. elections readinessBoth involve Anthropic, Claude
- NewsHow Claude's text watermarking worksBoth involve Anthropic, Claude
- NewsAttackers Crash Claude's 'Office Party' Flaw to Run Malicious Code on LinuxBoth involve Anthropic, Claude
- NewsClaude's new constitutionBoth involve Anthropic, Claude
- NewsDetecting and countering malicious uses of ClaudeBoth involve Anthropic, Claude
- NewsLLMs are making formal software verification feasible at a much larger scaleBoth involve Anthropic, Claude
- NewsMeasurements for understanding the pace of AI development inside frontier labsBoth involve Anthropic, Claude
- NewsWhat 81,000 people told us about AI economicsBoth involve Anthropic, Claude
- NewsDetecting and countering misuse of AI: August 2025Both involve Anthropic, Claude
- NewsStatement on the comments from Secretary of War Pete HegsethBoth involve Anthropic, Claude
- NewsAn alignment assessment of recent cybersecurity incidentsBoth involve Anthropic, Claude Opus 4.7
- NewsEnabling independent research on how people use ClaudeBoth involve Anthropic, Claude
- NewsThoughts on America’s AI Action PlanBoth involve Anthropic, Claude Opus 4.7
- NewsMeasuring LLMs’ ability to develop exploitsBoth involve Anthropic, Claude Opus 4.7
About
- Anthropic
- Claude