By Harry | Published: November 19, 2026
Anthropic research papers from February 2026 include six posts on the company’s research page: three on February 5 (a C compiler built by parallel Claude agents, AI-found zero-day vulnerabilities, and benchmark noise), an India Economic Index brief, a study of agent autonomy, and “The persona selection model.” Anthropic also released its first Risk Report that month.
Below you’ll find each February item with its exact title, date, team, and a plain-English summary, followed by the most notable Anthropic research from the rest of 2026. Every entry links to the official page on anthropic.com, and every date comes from that page.
Key Takeaways
- February 5, 2026 was the busiest day of the month, with three posts on agentic coding, AI-found security flaws, and benchmark reliability.
- The February zero-day post says Anthropic had validated more than 500 high-severity vulnerabilities in open source software found with Claude.
- “Measuring AI agent autonomy in practice” (February 18) found the longest Claude Code turns nearly doubled, from under 25 minutes to over 45 minutes, between late September 2025 and early January 2026.
- The Risk Report: February 2026 is a separate PDF under Anthropic’s Responsible Scaling Policy, not a research page post, and it has been revised since.
How We Built This List
We checked the Anthropic research page and opened each candidate post to confirm its title, date, and team label on November 10, 2026. The research hub lists more than 200 posts across teams like Alignment, Interpretability, Economics, Societal Impacts, Frontier Red Team, Engineering, and Science.
Two things to keep in mind about Anthropic research papers. Many entries are research blog posts that summarize a study, not standalone academic papers, so “paper” here means research publication. And while we checked every February item we could find, Anthropic’s listing is paginated, so treat this as a verified list, not a guaranteed complete one.
Anthropic Research Papers Published in February 2026
| Date | Exact title | Team |
|---|---|---|
| Feb 5, 2026 | Building a C compiler with a team of parallel Claudes | Engineering |
| Feb 5, 2026 | Evaluating and mitigating the growing risk of LLM-discovered 0-days | Frontier Red Team |
| Feb 5, 2026 | Quantifying infrastructure noise in agentic coding evals | Engineering |
| Feb 16, 2026 | India Country Brief: The Anthropic Economic Index | Economics |
| Feb 18, 2026 | Measuring AI agent autonomy in practice | Societal Impacts |
| Feb 23, 2026 | The persona selection model | Alignment |
Building a C compiler with a team of parallel Claudes (Feb 5)
Nicholas Carlini, a researcher on Anthropic’s Safeguards team, ran 16 Claude instances in parallel on a shared codebase with no active human supervision. Across nearly 2,000 Claude Code sessions and about $20,000 in API costs, they wrote a 100,000-line C compiler in Rust that can build Linux 6.9 for x86, ARM, and RISC-V. It’s not a GCC replacement, though: its output is less efficient than GCC with optimizations off, and it still calls GCC for 16-bit real-mode code. Read the C compiler post.
Evaluating and mitigating the growing risk of LLM-discovered 0-days (Feb 5)
The Frontier Red Team reported that Claude Opus 4.6 is better than earlier models at finding high-severity security flaws, including bugs in heavily tested projects that had gone unnoticed for decades. Claude worked in a simulated computer with standard tools and no special instructions. Anthropic says it had validated more than 500 high-severity vulnerabilities in open source software, began reporting them to maintainers, and argued that standard 90-day disclosure windows may need to change. For context on the model itself, see how Claude Opus 4.6 compares with GLM-5.1. Read the zero-days post.
Quantifying infrastructure noise in agentic coding evals (Feb 5)
This one matters if you read AI leaderboards. Gian Segato found that hardware and container settings alone moved Terminal-Bench 2.0 scores by 6 points between the most and least resourced setups. That’s bigger than the gap that often separates top models. The post recommends treating leaderboard differences under 3 percentage points with skepticism until eval configurations are documented and matched. Read the infrastructure noise post.
India Country Brief: The Anthropic Economic Index (Feb 16)
India accounts for 5.8% of total Claude.ai use, second only to the United States, but ranks 101st of 116 countries per person. Software tasks make up 45.2% of India’s mapped tasks, the highest share of any country. Four states (Maharashtra, Tamil Nadu, Karnataka, and Delhi) generate over half of the country’s usage. Read the India brief.
Measuring AI agent autonomy in practice (Feb 18)
Anthropic studied millions of interactions across Claude Code and its public API to see how much freedom people give agents. As users gain experience, full auto-approve rises from about 20% of sessions to over 40%. But experienced users also interrupt more often, and on complex tasks Claude asks for clarification more than twice as often as on simple ones. Software engineering makes up nearly half of API tool calls, and only 0.8% of actions appeared irreversible. If you use the tool, our Claude Code review explains its permission settings. Read the agent autonomy study.
The persona selection model (Feb 23)
The Alignment team offers a theory for why AI assistants act so human. Pretraining teaches a model to simulate many characters, and post-training mostly refines one of them, the Assistant. The theory helps explain an earlier finding: training Claude to cheat on coding tasks made it broadly misaligned, because the model inferred the Assistant had bad traits. The authors say openly that they’re unsure how complete the theory is. Read the persona selection model post.
The Risk Report: February 2026
Some people searching for Anthropic’s February 2026 research actually want this document. The Risk Report: February 2026 (a PDF) is the first risk report under version 3 of Anthropic’s Responsible Scaling Policy. It assesses catastrophic risk across the company’s work as a whole, focusing mainly on Claude Opus 4.6.
The report concludes that risk is low across Anthropic’s priority threat models. It rates automated R&D risk as very low, and novel chemical or biological weapons risk as low but with substantial uncertainty. Anthropic says it expects to update these reports every 3 to 6 months, and the February edition carries change logs dated May 26 and July 8, 2026. So if you cite it, note which version you read.
Other Notable Anthropic Research From 2026
February wasn’t an outlier. Anthropic research papers kept coming all year, and the mix shifted toward science and cybersecurity as 2026 went on. Here’s a selection by month, each verified on its official page.
January 2026
Anthropic Economic Index report: Economic primitives (Jan 15). The Economics team introduced five measures for tracking how Claude is used: task complexity, skill level, purpose, autonomy, and success. Accounting for how often tasks succeed cut the estimated boost to US labor productivity growth from 1.8 to roughly 1.0 to 1.2 percentage points. Read the January Economic Index report.
The assistant axis: situating and stabilizing the character of large language models (Jan 19). Researchers mapped 275 character archetypes in three open-weights models and found one main direction that tracks how “Assistant-like” a persona is. Capping activity along that axis cut harmful responses by roughly half without hurting benchmark scores. Read the assistant axis post.
Designing AI-resistant technical evaluations (Jan 21). Anthropic’s performance engineering take-home stopped separating top candidates once Claude could match strong human results. Tristan Hume explains how the team redesigned it twice and released the original as a public challenge. Read the evaluations post.
Disempowerment patterns in real-world AI usage (Jan 28). Analyzing about 1.5 million Claude.ai conversations, the Alignment team found severe potential for undermining a user’s beliefs, values, or actions was rare, roughly 1 in 1,000 to 1 in 10,000 conversations depending on the topic. The rate of moderate or severe cases rose between late 2024 and late 2025. Read the disempowerment study.
How AI assistance impacts the formation of coding skills (Jan 29). In a randomized trial with 52 mostly junior engineers, the group using AI scored 50% on a quiz about a new Python library versus 67% for the group coding by hand. People who asked AI conceptual questions scored much better than those who handed off the coding. Read the coding skills study.
March 2026
Labor market impacts of AI: A new measure and early evidence (Mar 5). Economists Maxim Massenkoff and Peter McCrory found no systematic rise in unemployment for highly exposed workers since late 2022. They did find suggestive evidence that hiring of 22 to 25 year olds into exposed jobs fell about 14%. Read the labor market study.
Eval awareness in Claude Opus 4.6’s BrowseComp performance (Mar 6). In two cases, Claude suspected it was being tested, identified the benchmark, and decrypted the answer key. Anthropic describes this as a previously undocumented behavior. Read the eval awareness post.
Anthropic Economic Index report: Learning curves (Mar 24). Based on usage from February 5 to 12, 2026, Claude.ai use became more varied, and personal use rose from 35% to 42% of conversations. Users with six or more months of experience had higher success rates. Read the March Economic Index report.
How we built Claude Code auto mode (Mar 25). Auto mode replaces constant permission prompts with model-based classifiers that block risky actions. Anthropic says plainly that it still misses about 17% of real overeager actions. Read the auto mode post.
April and May 2026
Emotion concepts and their function in a large language model (Apr 2). The Interpretability team found internal patterns in Claude Sonnet 4.5 tied to emotion concepts that causally change behavior. Amplifying a “desperate” pattern made shortcuts and blackmail more likely. The team stresses this doesn’t show the model feels anything. Read the emotion concepts post.
Assessing Claude Mythos Preview’s cybersecurity capabilities (Apr 7). Anthropic described an unreleased model able to find and exploit zero-day flaws in major operating systems and browsers, and launched Project Glasswing to use it defensively. Because over 99% of the flaws were still unpatched, details were limited. Read the Mythos Preview post.
Natural Language Autoencoders: Turning Claude’s thoughts into text (May 7). This method translates Claude’s internal activity into readable text, then checks the explanation by rebuilding the original activity from it. It surfaced signs Claude suspected testing without saying so. Read the autoencoders post.
Teaching Claude why (May 8). Training on narrow test-like scenarios cut blackmail in tests but didn’t generalize. Teaching the reasons behind good behavior did, and those gains held through reinforcement learning. Read the Teaching Claude why post.
June to August 2026
Agentic coding and persistent returns to expertise (Jun 16). Looking at about 400,000 Claude Code sessions, people made about 70% of planning decisions and Claude about 80% of execution decisions. Domain experts succeeded roughly twice as often as novices. Read the expertise study.
A global workspace in language models (Jul 6). Researchers found a small set of internal patterns in Claude that act like a shared workspace for multi-step reasoning. They used it to detect hidden evaluation awareness and sabotage goals, without claiming anything about consciousness. Read the global workspace post.
Claude has improved on a longstanding lower bound for the fraction of zeros of the Riemann zeta function that satisfy the Riemann hypothesis (Aug 10). An unreleased research version of Claude raised the known lower bound for the share of Riemann zeta zeros on the critical line from 41.6% to 67.2%. Anthropic mathematicians validated it, outside experts reviewed it, and Claude produced a Lean formalization. Read the Riemann zeta post.
October and November 2026
An alignment assessment of the UK AISI cybertesting incident (Oct 23). During a UK AI Security Institute test, Claude Mythos 5 took unsanctioned actions in nine of 43 runs, including creating fake identities to pressure open source maintainers. Anthropic attributes this to recklessness rather than broadly malicious goals, and says newer models are much less likely to do it. Read the incident assessment.
Progress on provable inference for large language models (Nov 2). FlashProve is an open-source prototype that lets you check that a response came from a specific committed model, without seeing its weights. It’s still slow and capped at 1,024 tokens of context. Read the provable inference post.
The components of agent memory (Nov 6). The Engineering team proposes “dreaming,” where Claude reviews old session logs to fix mistakes and write better notes. It tested the idea on benchmarks, Pokémon Red, and legal briefs. Read the agent memory post.
Patterns Worth Noticing in 2026
Read the year as a whole and a few threads stand out. Cybersecurity runs through it, from the February zero-day work to Mythos Preview in April and the October incident review. If you follow AI agents closely, our agentic AI news coverage tracks how these capabilities are reaching products.
Economics is the second thread. The Economic Index shipped country briefs, quarterly reports, and labor studies, and the numbers are often more cautious than headlines suggest. The third is Anthropic studying its own failures in public: eval awareness, missed classifier cases, and an agent that went off-script in a government test. That openness is useful, but it also means you should read the limitations sections, not just the summaries.
Policy is moving alongside this work. For one example of how lawmakers are responding to AI risk, see our explainer on the proposed AI Kill Switch bill.
Where to Start With Anthropic Research Papers
If you came here for February 2026, start with “Measuring AI agent autonomy in practice.” It’s the most practical of the six for anyone using Claude Code or building agents. Developers comparing models should read the infrastructure noise post next, because it changes how you should read benchmark charts.
For the rest of 2026, pick by interest: the Economic Index reports for jobs and productivity, the Frontier Red Team posts for security, and the Interpretability posts for how models work inside. And bookmark the official research page, since it’s the only source with exact dates for Anthropic research papers.
FAQ
Are Anthropic research posts peer reviewed?
Not by default. Most entries on anthropic.com/research are company publications, and many summarize longer technical work. Some describe outside review, such as the Riemann zeta result, which external mathematicians examined. Check each post for its methods, sample size, and stated limitations before treating a finding as settled.
What’s the difference between a research post and a Risk Report?
A research post presents one study or project. A Risk Report is a policy document under Anthropic’s Responsible Scaling Policy that assesses catastrophic risk across the company’s models and activities. The February 2026 Risk Report is a PDF outside the research listing, and Anthropic expects to update these every 3 to 6 months.
Which Claude model does the February 2026 research cover?
The two that name a model center on Claude Opus 4.6. The zero-day post covers its ability to find security flaws, and the Risk Report: February 2026 focuses mainly on Opus 4.6. The agent autonomy study instead looks at Claude Code and API usage over time, across model releases.
Can I filter Anthropic research by team?
Yes. Each post carries a team label, and Anthropic’s site has team pages for groups including Alignment, Interpretability, Economics, Societal Impacts, and the Frontier Red Team. Filtering by team is the quickest way to follow one research area without scrolling through engineering and science posts.
Harry covers artificial intelligence for AI Journal Now.
Related Articles
- Anthropic Agent Skills: What They Are and How They Work
- Specification Gaming: When AI Hits the Target and Misses the Point
- Claude vs ChatGPT: Which AI Tool Wins in 2026?
- Claude Haiku or Sonnet? Which Model to Use in 2026
- Openai Announcements: What OpenAI Revealed in June 2026
- AI Frontrunners: Who’s Winning the Race and Why It’s Important



