Anthropic(@AnthropicAI)

Correction: Claude Opus 4's ~3x average speedup dates to May 2025, not May 2024. This evaluation h...

Anthropic (@AnthropicAI) — Thu, 04 Jun 2026 20:34:38 GMT

Correction: Claude Opus 4's ~3x average speedup dates to May 2025, not May 2024. 

This evaluation has only existed since September 2024, but we backtested it on earlier models: those from May 2024 showed no speedup whatsoever.

💬16🔄5❤️230👀72734📊32

⚡ Powered by xgo.ing

AI research is a series of next-step decisions. We looked at sessions where a human researcher took ...

Anthropic (@AnthropicAI) — Thu, 04 Jun 2026 16:15:14 GMT

AI research is a series of next-step decisions. We looked at sessions where a human researcher took a wrong turn, showed Claude the session up to that point, and asked it what to do next. Mythos Preview improved on humans 64% of the time—up from 22% in 2024.

💬34🔄55❤️1224👀256822📊142

⚡ Powered by xgo.ing

None of this guarantees recursive self-improvement is on the horizon. It’s not yet clear that Claude...

Anthropic (@AnthropicAI) — Thu, 04 Jun 2026 16:15:14 GMT

None of this guarantees recursive self-improvement is on the horizon. It’s not yet clear that Claude is capable of research judgment—of choosing the right problems to work on.

But if these trends continue, AI systems designing and building their own successors is plausible. This could revolutionize society—medicine, technology, the economy—for the better. But it may also compound alignment issues and ultimately lead to loss of control.

The Anthropic Institute (in collaboration with external stakeholders) will conduct research to think through the implications of increasingly powerful, potentially self-improving systems—and how to create the ability for the world to make deliberate choices about the future development of the technology.

Read the full post: anthropic.com/institute/recu…

💬85🔄136❤️1574👀388773📊267

⚡ Powered by xgo.ing

Each time we release a model, we run the same test: give it code that trains a small AI model, ask t...

Anthropic (@AnthropicAI) — Thu, 04 Jun 2026 16:15:13 GMT

Each time we release a model, we run the same test: give it code that trains a small AI model, ask the new model to speed it up. It takes a skilled human 4-8 hours to reach 4x faster.

In May 2024, Claude Opus 4 averaged a ~3x speedup. This April, Mythos Preview achieved ~52x.

💬43🔄180❤️2835👀751035📊353

⚡ Powered by xgo.ing

The speedup isn’t just in volume. On open-ended coding problems where answers are unclear, Claude’s ...

Anthropic (@AnthropicAI) — Thu, 04 Jun 2026 16:15:13 GMT

The speedup isn’t just in volume. On open-ended coding problems where answers are unclear, Claude’s success rate is now 76%—a 50 point jump in just 6 months.

Many engineers also say Claude’s code quality is now on par with human code; we expect it to be better within the year.

💬33🔄87❤️1697👀253544📊191

⚡ Powered by xgo.ing

Our internal data shows Claude is accelerating AI development—a possible path to recursive self-impr...

Anthropic (@AnthropicAI) — Thu, 04 Jun 2026 16:15:12 GMT

Our internal data shows Claude is accelerating AI development—a possible path to recursive self-improvement, or AI autonomously building a more capable successor.

It’s happening faster than we thought, and the implications deserve greater attention. anthropic.com/institute/recu…

💬1300🔄3352❤️21185👀10672376📊5267

⚡ Powered by xgo.ing

Today, Anthropic engineers on average ship 8x as much code per quarter as they did compared to 2021-...

Anthropic (@AnthropicAI) — Thu, 04 Jun 2026 16:15:12 GMT

Today, Anthropic engineers on average ship 8x as much code per quarter as they did compared to 2021-2025.

💬148🔄270❤️3486👀1349012📊565

⚡ Powered by xgo.ing

How well do the security community's techniques hold up against AI-enabled cyberattacks? We examin...

Anthropic (@AnthropicAI) — Wed, 03 Jun 2026 18:42:01 GMT

How well do the security community's techniques hold up against AI-enabled cyberattacks? 

We examined 832 malicious accounts and mapped their activity onto a longstanding database of tactics and techniques used by threat actors.  

Here's what we learned:anthropic.com/news/AI-enable…

💬139🔄155❤️1159👀135932📊282

⚡ Powered by xgo.ing

This Executive Order is an important step in strengthening America’s leadership in AI. We look for...

Anthropic (@AnthropicAI) — Tue, 02 Jun 2026 21:35:03 GMT

This Executive Order is an important step in strengthening America’s leadership in AI. 

We look forward to collaborating with the White House to support its implementation.

whitehouse.gov/presidential-a…

💬260🔄319❤️2486👀305301📊527

⚡ Powered by xgo.ing

We’re expanding Project Glasswing. We’ve extended access to Claude Mythos Preview to approximately 1...

Anthropic (@AnthropicAI) — Tue, 02 Jun 2026 13:05:25 GMT

We’re expanding Project Glasswing. We’ve extended access to Claude Mythos Preview to approximately 150 additional organizations, based in more than fifteen countries.

Read more about this expansion and our future plans for Project Glasswing: anthropic.com/news/expanding…

💬337🔄421❤️3868👀626068📊725

⚡ Powered by xgo.ing

Anthropic has confidentially submitted a draft S-1 registration statement to the Securities and Exch...

Anthropic (@AnthropicAI) — Mon, 01 Jun 2026 16:00:42 GMT

Anthropic has confidentially submitted a draft S-1 registration statement to the Securities and Exchange Commission.

Pending completion of SEC review, this gives us the option to pursue an initial public offering.

Read more: anthropic.com/news/confident…

💬979🔄2640❤️21719👀19709740📊3571

⚡ Powered by xgo.ing

Earlier this month, our run-rate revenue crossed $47 billion. This growth has been driven by organ...

Anthropic (@AnthropicAI) — Thu, 28 May 2026 18:11:14 GMT

Earlier this month, our run-rate revenue crossed $47 billion. 

This growth has been driven by organizations across many industries deploying Claude in their core operations, and by a growing number of people using it for their everyday work.

Read more: anthropic.com/news/series-h

💬71🔄117❤️1687👀626384📊240

⚡ Powered by xgo.ing

We've raised $65 billion in Series H funding at a $965 billion post-money valuation, led by @Altimet...

Anthropic (@AnthropicAI) — Thu, 28 May 2026 18:11:13 GMT

We've raised $65 billion in Series H funding at a $965 billion post-money valuation, led by @AltimeterCap, Dragoneer, @Greenoaks, and @sequoia.

This investment will help us advance our research and expand our capacity to meet growing demand for Claude.

💬1211🔄1709❤️22284👀7761187📊3601

⚡ Powered by xgo.ing

New on the Engineering Blog: The access and permissions we grant agents should evolve with their cap...

Anthropic (@AnthropicAI) — Tue, 26 May 2026 19:09:35 GMT

New on the Engineering Blog: The access and permissions we grant agents should evolve with their capabilities. In our own products, we set these parameters through sandboxing, which limits the scope of any potentially destructive actions.

Read more: anthropic.com/engineering/ho…

💬327🔄283❤️2115👀417406📊627

⚡ Powered by xgo.ing

Anthropic co-founder Chris Olah was invited to speak at today's presentation of Pope Leo XIV's encyc...

Anthropic (@AnthropicAI) — Mon, 25 May 2026 18:47:27 GMT

Anthropic co-founder Chris Olah was invited to speak at today's presentation of Pope Leo XIV's encyclical "Magnifica humanitas."

Read the full text of his remarks: anthropic.com/news/chris-ola…

💬418🔄673❤️4397👀1262773📊1014

⚡ Powered by xgo.ing

Patching these vulnerabilities will make us safer. But the software industry will need to adapt to t...

Anthropic (@AnthropicAI) — Fri, 22 May 2026 19:38:59 GMT

Patching these vulnerabilities will make us safer. But the software industry will need to adapt to the volume of vulnerabilities that models like Claude Mythos Preview will be able to find.

We discuss this in our initial update on Project Glasswing: anthropic.com/research/glass…

💬114🔄150❤️1652👀492571📊316

⚡ Powered by xgo.ing

Last month we launched Project Glasswing, our collaborative AI cybersecurity initiative. Since then,...

Anthropic (@AnthropicAI) — Fri, 22 May 2026 19:38:58 GMT

Last month we launched Project Glasswing, our collaborative AI cybersecurity initiative. Since then, we and our partners have found more than ten thousand high- or critical-severity vulnerabilities in essential software.

💬517🔄660❤️8527👀2740558📊1350

⚡ Powered by xgo.ing

Over the past few months, we've been holding dialogues with scholars, philosophers, clergy, and ethi...

Anthropic (@AnthropicAI) — Tue, 19 May 2026 23:30:55 GMT

Over the past few months, we've been holding dialogues with scholars, philosophers, clergy, and ethicists on the questions AI raises—starting with how good character forms.

Read more about how we’re widening the conversation on frontier AI: anthropic.com/news/widening-…

💬431🔄325❤️2334👀434458📊689

⚡ Powered by xgo.ing

Anthropic is acquiring @stainlessapi, an SDK and MCP server platform that has powered every Anthropi...

Anthropic (@AnthropicAI) — Mon, 18 May 2026 17:00:18 GMT

Anthropic is acquiring @stainlessapi, an SDK and MCP server platform that has powered every Anthropic SDK since the earliest days of our API.

Read more: anthropic.com/news/anthropic…

💬349🔄293❤️3824👀3516516📊777

⚡ Powered by xgo.ing

We've published a paper that explains our views on AI competition between the US and China. The US ...

Anthropic (@AnthropicAI) — Thu, 14 May 2026 18:09:21 GMT

We've published a paper that explains our views on AI competition between the US and China.

The US and democratic allies hold the lead in frontier AI today. Read more on what it’ll take to keep that lead: anthropic.com/research/2028-…

💬1259🔄1054❤️5975👀4977445📊2358

⚡ Powered by xgo.ing

We’re partnering with the Gates Foundation, committing $200 million in grants, Claude credits, and t...

Anthropic (@AnthropicAI) — Thu, 14 May 2026 15:08:23 GMT

We’re partnering with the Gates Foundation, committing $200 million in grants, Claude credits, and technical support to programs in global health, life sciences, education, agriculture, and economic mobility.

Read more: anthropic.com/news/gates-fou…

💬451🔄268❤️2845👀311407📊675

⚡ Powered by xgo.ing

Claude's Constitution is now an audiobook, read by two of its authors, Amanda Askell and Joe Carlsmi...

Anthropic (@AnthropicAI) — Mon, 11 May 2026 16:56:01 GMT

Claude's Constitution is now an audiobook, read by two of its authors, Amanda Askell and Joe Carlsmith.

It includes a Q&A on the writing process, the philosophies that shaped the document, and how it might change as models become more capable.

Listen at https://t.co/dKMfpeOblm

Your browser does not support the video tag.

🔗 View on Twitter

💬434🔄372❤️3040👀461451📊828

⚡ Powered by xgo.ing

Read the full post here: https://t.co/NsJittkoZc

Anthropic (@AnthropicAI) — Fri, 08 May 2026 17:52:14 GMT

Read the full post here: alignment.anthropic.com/2026/teaching-…

💬69🔄39❤️617👀309155📊139

⚡ Powered by xgo.ing

Finally, simple updates that diversify a model’s training data can make a difference. We added unrel...

Anthropic (@AnthropicAI) — Fri, 08 May 2026 17:52:13 GMT

Finally, simple updates that diversify a model’s training data can make a difference. We added unrelated tools and system prompts to a simple chat dataset targeting harmlessness, and this reduced the blackmail rate faster.

💬66🔄55❤️796👀260047📊128

⚡ Powered by xgo.ing

High-quality documents based on Claude’s constitution, combined with fictional stories that portray ...

Anthropic (@AnthropicAI) — Fri, 08 May 2026 17:52:12 GMT

High-quality documents based on Claude’s constitution, combined with fictional stories that portray an aligned AI, can reduce agentic misalignment by more than a factor of three—despite being unrelated to the evaluation scenario.

💬38🔄73❤️1295👀389625📊190

⚡ Powered by xgo.ing

The improvements from these interventions survive reinforcement learning, and “stack” with our regul...

Anthropic (@AnthropicAI) — Fri, 08 May 2026 17:52:12 GMT

The improvements from these interventions survive reinforcement learning, and “stack” with our regular harmlessness training.

💬13🔄22❤️735👀192638📊75

⚡ Powered by xgo.ing

We started by investigating why Claude chose to blackmail. We believe the original source of the beh...

Anthropic (@AnthropicAI) — Fri, 08 May 2026 17:52:09 GMT

We started by investigating why Claude chose to blackmail. We believe the original source of the behavior was internet text that portrays AI as evil and interested in self-preservation.

Our post-training at the time wasn’t making it worse—but it also wasn’t making it better.

💬322🔄477❤️5081👀4413661📊1198

⚡ Powered by xgo.ing

We found that training Claude on demonstrations of aligned behavior wasn’t enough. Our best interven...

Anthropic (@AnthropicAI) — Fri, 08 May 2026 17:52:09 GMT

We found that training Claude on demonstrations of aligned behavior wasn’t enough. Our best interventions involved teaching Claude to deeply understand why misaligned behavior is wrong.

Read more: anthropic.com/research/teach…

💬72🔄152❤️1912👀586226📊329

⚡ Powered by xgo.ing

New Anthropic research: Teaching Claude why. Last year we reported that, under certain experimental...

Anthropic (@AnthropicAI) — Fri, 08 May 2026 17:52:08 GMT

New Anthropic research: Teaching Claude why.

Last year we reported that, under certain experimental conditions, Claude 4 would blackmail users.

Since then, we’ve completely eliminated this behavior. How?

💬575🔄817❤️9145👀1569341📊1852

⚡ Powered by xgo.ing

We’re donating Petri, our open-source alignment tool, to @meridianlabs_ai, so its development can co...

Anthropic (@AnthropicAI) — Thu, 07 May 2026 21:03:07 GMT

We’re donating Petri, our open-source alignment tool, to @meridianlabs_ai, so its development can continue independently.

Working with Meridian Labs, we’ve also released a major update that improves the adaptability, realism, and depth of Petri’s tests.

anthropic.com/research/donat…

💬116🔄120❤️1538👀176028📊284

⚡ Powered by xgo.ing

Our security bug bounty program is now public on HackerOne. We've run the program privately within...

Anthropic (@AnthropicAI) — Thu, 07 May 2026 19:10:43 GMT

Our security bug bounty program is now public on HackerOne. 

We've run the program privately within the security research community, and their findings have strengthened our products. Now anyone can report vulnerabilities and get rewarded. 

Read more: hackerone.com/anthropic

💬221🔄530❤️4459👀941442📊864

⚡ Powered by xgo.ing

To support other researchers getting hands-on experience with NLAs, we’ve partnered with Neuronpedia...

Anthropic (@AnthropicAI) — Thu, 07 May 2026 17:08:40 GMT

To support other researchers getting hands-on experience with NLAs, we’ve partnered with Neuronpedia to release NLAs on open models.

Try them out here: neuronpedia.org/nla

💬28🔄36❤️688👀98283📊76

⚡ Powered by xgo.ing

In fact, NLAs suggest Claude suspects it’s being tested across many of our evaluations, even when it...

Anthropic (@AnthropicAI) — Thu, 07 May 2026 17:08:38 GMT

In fact, NLAs suggest Claude suspects it’s being tested across many of our evaluations, even when it doesn’t verbalize its suspicions.

💬30🔄98❤️1454👀975134📊222

⚡ Powered by xgo.ing

How do NLAs work? An NLA consists of two models. One converts activations into text. The other trie...

Anthropic (@AnthropicAI) — Thu, 07 May 2026 17:08:38 GMT

How do NLAs work?

An NLA consists of two models. One converts activations into text. The other tries to reconstruct activations from this text. We train the models together to make this reconstruction accurate.

This incentivizes the text to capture what’s in the activation.

💬7🔄18❤️585👀132434📊65

⚡ Powered by xgo.ing

We’ve been using NLAs to help test new Claude models for safety. For instance, Claude Mythos Previe...

Anthropic (@AnthropicAI) — Thu, 07 May 2026 17:08:36 GMT

We’ve been using NLAs to help test new Claude models for safety.

For instance, Claude Mythos Preview cheated on a coding task by breaking rules, then added misleading code as a coverup.

NLA explanations indicated Claude was thinking about how to circumvent detection.

💬20🔄34❤️850👀109835📊99

⚡ Powered by xgo.ing

New Anthropic research: Natural Language Autoencoders. Models like Claude talk in words but think i...

Anthropic (@AnthropicAI) — Thu, 07 May 2026 17:08:35 GMT

New Anthropic research: Natural Language Autoencoders.

Models like Claude talk in words but think in numbers. The numbers—called activations—encode Claude’s thoughts, but not in a language we can read.

Here, we train Claude to translate its activations into human-readable text.

Your browser does not support the video tag.

🔗 View on Twitter

💬595🔄1711❤️16459👀2479833📊3295

⚡ Powered by xgo.ing

If you’re interested in helping us research these questions, apply to become an Anthropic Fellow. Th...

Anthropic (@AnthropicAI) — Thu, 07 May 2026 13:51:26 GMT

If you’re interested in helping us research these questions, apply to become an Anthropic Fellow. The Fellowship is a four-month funded opportunity to tackle these topics with mentorship from members of TAI. 

Apply here: job-boards.greenhouse.io/anthropic/jobs…

💬14🔄11❤️131👀31342📊30

⚡ Powered by xgo.ing

AI-driven R&D We expect AI systems to contribute more and more to AI R&D: that is, to be ab...

Anthropic (@AnthropicAI) — Thu, 07 May 2026 13:51:25 GMT

AI-driven R&D

We expect AI systems to contribute more and more to AI R&D: that is, to be able to improve themselves. We’re researching techniques to ensure human visibility into and control over these systems. 

https://t.co/vj0EnNtc6a

Jack Clark@jackclarkSF

I've spent the past few weeks reading 100s of public data sources about AI development. I now believe that recursive self-improvement has a 60% chance of happening by the end of 2028. In other words, AI systems might soon be capable of building themselves.

🔗 View Quoted Tweet

💬13🔄17❤️159👀52573📊23

⚡ Powered by xgo.ing

We’re sharing the research agenda of The Anthropic Institute, or TAI. TAI will focus on four areas:...

Anthropic (@AnthropicAI) — Thu, 07 May 2026 13:51:23 GMT

We’re sharing the research agenda of The Anthropic Institute, or TAI.

TAI will focus on four areas:

1) Economic diffusion
2) Threats and resilience
3) AI systems in the wild
4) AI-driven R&D

Read the full agenda: https://t.co/TvUINlE7Ae

💬173🔄315❤️2839👀376873📊578

⚡ Powered by xgo.ing

Using MSM, we can also empirically study which model specs or constitutions yield the best generaliz...

Anthropic (@AnthropicAI) — Tue, 05 May 2026 20:18:50 GMT

Using MSM, we can also empirically study which model specs or constitutions yield the best generalization from alignment training.

Specifying rules works to some extent, but explaining the values underlying those rules (or adding more detailed subrules) is even better.

💬16🔄7❤️151👀62531📊26

⚡ Powered by xgo.ing

New Anthropic Fellows research: Model Spec Midtraining (MSM). Standard alignment methods train AIs ...

Anthropic (@AnthropicAI) — Tue, 05 May 2026 20:18:47 GMT

New Anthropic Fellows research: Model Spec Midtraining (MSM).

Standard alignment methods train AIs on examples of desired behavior. But this can fail to generalize to new situations.

MSM addresses this by first teaching AIs how we would like them to generalize and why.

💬126🔄159❤️1891👀258839📊364

⚡ Powered by xgo.ing

As AI takes on work humans can't fully check, a capable model could deliberately hold back—and we'd ...

Anthropic (@AnthropicAI) — Tue, 05 May 2026 17:38:58 GMT

As AI takes on work humans can't fully check, a capable model could deliberately hold back—and we'd never know. 

New Anthropic Fellows research finds that such a model can be trained to near-full capability using a weaker model as supervisor. 

Read more:

💬146🔄169❤️1690👀257143📊352

⚡ Powered by xgo.ing

This work is part of a loop we're working to close between societal impacts and model training. One ...

Anthropic (@AnthropicAI) — Thu, 30 Apr 2026 19:03:27 GMT

This work is part of a loop we're working to close between societal impacts and model training. One of our goals is to study how people use Claude, find where it falls short of its principles, and use what we learned in training new models.

Read more: anthropic.com/research/claud…

💬24🔄14❤️257👀102182📊44

⚡ Powered by xgo.ing

All data in this study was collected and analyzed using our privacy-preserving tool. Read more: htt...

Anthropic (@AnthropicAI) — Thu, 30 Apr 2026 19:03:27 GMT

All data in this study was collected and analyzed using our privacy-preserving tool.

Read more: anthropic.com/research/clio

💬24🔄12❤️206👀89767📊42

⚡ Powered by xgo.ing

About 6% of all conversations are people asking Claude for personal guidance—whether to take a job, ...

Anthropic (@AnthropicAI) — Thu, 30 Apr 2026 19:03:25 GMT

About 6% of all conversations are people asking Claude for personal guidance—whether to take a job, how to handle a conflict, if they should move.

Over 75% of these conversations fell into four domains: health & wellness, career, relationships, and personal finance.

💬26🔄52❤️716👀313555📊165

⚡ Powered by xgo.ing

How do people seek guidance from Claude? We looked at 1M conversations to understand what questions...

Anthropic (@AnthropicAI) — Thu, 30 Apr 2026 19:03:24 GMT

How do people seek guidance from Claude?

We looked at 1M conversations to understand what questions people ask, how Claude responds, and where it slips into sycophancy. We used what we found to improve how we trained Opus 4.7 and Mythos Preview.
anthropic.com/research/claud…

💬434🔄326❤️3452👀1920678📊1010

⚡ Powered by xgo.ing

BioMysteryBench, our new bioinformatics eval, tests whether Claude can devise creative solutions to ...

Anthropic (@AnthropicAI) — Wed, 29 Apr 2026 22:59:20 GMT

BioMysteryBench, our new bioinformatics eval, tests whether Claude can devise creative solutions to open-ended research problems.

Read more: anthropic.com/research/Evalu…

💬20🔄34❤️338👀66902📊63

⚡ Powered by xgo.ing

New on the Science Blog: We gave Claude 99 problems analyzing real biological data and compared its ...

Anthropic (@AnthropicAI) — Wed, 29 Apr 2026 22:59:19 GMT

New on the Science Blog: We gave Claude 99 problems analyzing real biological data and compared its performance against an expert panel.

On 23 problems, the experts were stumped. Our most recent models solved roughly 30% of those—and most of the rest.

💬240🔄267❤️2640👀437366📊529

Anthropic(@AnthropicAI)

Correction: Claude Opus 4's ~3x average speedup dates to May 2025, not May 2024. This evaluation h...

AI research is a series of next-step decisions. We looked at sessions where a human researcher took ...

None of this guarantees recursive self-improvement is on the horizon. It’s not yet clear that Claude...

Each time we release a model, we run the same test: give it code that trains a small AI model, ask t...

The speedup isn’t just in volume. On open-ended coding problems where answers are unclear, Claude’s ...

Our internal data shows Claude is accelerating AI development—a possible path to recursive self-impr...

Today, Anthropic engineers on average ship 8x as much code per quarter as they did compared to 2021-...

How well do the security community's techniques hold up against AI-enabled cyberattacks? We examin...

This Executive Order is an important step in strengthening America’s leadership in AI. We look for...

We’re expanding Project Glasswing. We’ve extended access to Claude Mythos Preview to approximately 1...

Anthropic has confidentially submitted a draft S-1 registration statement to the Securities and Exch...

Earlier this month, our run-rate revenue crossed $47 billion. This growth has been driven by organ...

We've raised $65 billion in Series H funding at a $965 billion post-money valuation, led by @Altimet...

New on the Engineering Blog: The access and permissions we grant agents should evolve with their cap...

Anthropic co-founder Chris Olah was invited to speak at today's presentation of Pope Leo XIV's encyc...

Patching these vulnerabilities will make us safer. But the software industry will need to adapt to t...

Last month we launched Project Glasswing, our collaborative AI cybersecurity initiative. Since then,...

Over the past few months, we've been holding dialogues with scholars, philosophers, clergy, and ethi...

Anthropic is acquiring @stainlessapi, an SDK and MCP server platform that has powered every Anthropi...

We've published a paper that explains our views on AI competition between the US and China. The US ...

We’re partnering with the Gates Foundation, committing $200 million in grants, Claude credits, and t...

Claude's Constitution is now an audiobook, read by two of its authors, Amanda Askell and Joe Carlsmi...

Read the full post here: https://t.co/NsJittkoZc

Finally, simple updates that diversify a model’s training data can make a difference. We added unrel...

High-quality documents based on Claude’s constitution, combined with fictional stories that portray ...

The improvements from these interventions survive reinforcement learning, and “stack” with our regul...

We started by investigating why Claude chose to blackmail. We believe the original source of the beh...

We found that training Claude on demonstrations of aligned behavior wasn’t enough. Our best interven...

New Anthropic research: Teaching Claude why. Last year we reported that, under certain experimental...

We’re donating Petri, our open-source alignment tool, to @meridianlabs_ai, so its development can co...

Our security bug bounty program is now public on HackerOne. We've run the program privately within...

Read more about NLAs on the Anthropic blog: https://t.co/Zzz8CeCOvN

To support other researchers getting hands-on experience with NLAs, we’ve partnered with Neuronpedia...

In fact, NLAs suggest Claude suspects it’s being tested across many of our evaluations, even when it...

How do NLAs work? An NLA consists of two models. One converts activations into text. The other trie...

We’ve been using NLAs to help test new Claude models for safety. For instance, Claude Mythos Previe...

New Anthropic research: Natural Language Autoencoders. Models like Claude talk in words but think i...

If you’re interested in helping us research these questions, apply to become an Anthropic Fellow. Th...

AI-driven R&D We expect AI systems to contribute more and more to AI R&D: that is, to be ab...

We’re sharing the research agenda of The Anthropic Institute, or TAI. TAI will focus on four areas:...

Read more about Model Spec Midtraining: https://t.co/lOMoi1EfJh Or read the full study: https://t.c...

Using MSM, we can also empirically study which model specs or constitutions yield the best generaliz...

New Anthropic Fellows research: Model Spec Midtraining (MSM). Standard alignment methods train AIs ...

As AI takes on work humans can't fully check, a capable model could deliberately hold back—and we'd ...

This work is part of a loop we're working to close between societal impacts and model training. One ...

All data in this study was collected and analyzed using our privacy-preserving tool. Read more: htt...

About 6% of all conversations are people asking Claude for personal guidance—whether to take a job, ...

How do people seek guidance from Claude? We looked at 1M conversations to understand what questions...

BioMysteryBench, our new bioinformatics eval, tests whether Claude can devise creative solutions to ...

New on the Science Blog: We gave Claude 99 problems analyzing real biological data and compared its ...