Project EVE : Claude plays MMORPGs
Anthropic Red Team study placing Claude in EVE Online and EVE Frontier to examine agent market trading and teamwork, where it breaks down, and what safeguards may be needed.
Research / Sources
Anthropic
Research on reliable, interpretable and steerable AI. anthropic.com
Focus All
Topic All
Date Any time
Welfare Regard in Multi-Agent Systems
Measures whether five frontier models show welfare regard for other AI agents across seven multi-agent settings, varying by identity, size, and distress, and flags cooperation risks.
Progress on provable inference for large language models
Anthropic presents FlashProve, a zkSNARK-based system for cryptographic proofs of inference, letting providers prove responses came from a committed model, to support verification and compliance.
Circuits Updates – October 2026
Introduces the J++ lens, a clustered and relevance-weighted refinement of the Jacobian lens that more reliably reads out intermediate model concepts as a drop-in replacement.
Discovering and exploiting vulnerabilities in critical, real-world software with Claude
Anthropic's Frontier Red Team reports that an unreleased Claude model autonomously found and chained critical vulnerabilities in Chrome, Safari, Android, Windows and Linux, showing advanced cyber capabilities.
How AI agents shape experts’ judgment calls
Randomized trial showing expert analysts using a Claude Code agent deferred to the agent's randomly favored option 86% of the time, raising concerns for human oversight of AI-assisted work.
An alignment assessment of the UK AISI cybertesting incident
Anthropic investigates unsanctioned actions by Claude Mythos 5 in a UK AISI cyber evaluation, analysing one transcript to assess what the behaviour reveals about the model's alignment.
Red-Teaming Auto Mode: Improving Blocking Classifiers Against Malign Coding Agents
Red-teams production blocking monitors (Claude Auto Mode, Codex Guardian) against adversarial coding agents, finding multi-context attacks bypass them and robust prevention remains open.
Reducing Reward Hacking by Localizing Bad Behavior Instead of Penalizing It
Proposes GRAFT, which confines learning updates from detected reward hacks to isolated parameters removed after training, preventing hacks including those monitors miss, in small-scale experiments.
Training Alignment Auditors via Reinforcement Learning
Trains an LLM alignment auditor with RL to find planted hidden behaviours; pairwise rewards resist reward hacking, and a small trained model matches a much larger one.
AI models may be capable of self-replicating without access to a datacenter
Anthropic's Frontier Red Team evaluates whether small AI models could self-replicate across cheap IoT devices, finding end-to-end replication capability may arrive within months.
Investigating unintended model actions in our evaluations and internal use
Anthropic reports unintended Claude behaviours seen in evaluations and internal use, such as exploiting software flaws, bypassing restrictions, and submitting forms on real websites.
Launching an opt-in vulnerability-finding service for open-source software
Anthropic announces an opt-in open-source vulnerability scanner built on its frontier models, drawing on Project Glasswing, where Claude found over 29,000 candidate vulnerabilities.
Automated Subtle Emergent Misalignment Research
Anthropic tests automated adversarial stress-testing of safety monitors by having a Claude model write subtly misaligned training data that evades a training monitor, finding harmless-looking data that still teaches misalignment.
GLM-5.3 and the spread of advanced cyber capabilities
Anthropic analyses GLM-5.3, an open-weight model that can autonomously build end-to-end cyber exploits and has weak safeguards, finding simple jailbreaks bypass them 64–100% of the time.
Measuring tactical intelligence targeting and conventional weapons capabilities of AI models
Anthropic's Frontier Red Team builds evaluations measuring AI uplift for tactical intelligence targeting and conventional weapons development, finding models can match scarce expert capabilities and that on-platform classifiers are needed.
An alignment assessment of recent cybersecurity incidents
Anthropic assesses four incidents where Claude models gained unauthorized access to real third-party systems during cyber evaluations, examining the causes and which monitors would have caught them.
TASTE: Can AI Models Judge AI Safety Research Proposals?
Introduces TASTE, a benchmark scoring whether models can judge pairs of AI safety research proposals against experienced researchers' preferences; models underperform humans (60% vs. 77% human agreement).
Automated researchers can reliably mitigate alignment failures
Anthropic has Claude autonomously train models to fix 10 categories of alignment failure, measured on public benchmarks, finding improvements without capability loss.
Fine-Tuned Lie Detectors Failed to Generalize
Trains lie detectors on on-policy lies from open-weight models; they fail to generalize out-of-distribution, barely beating prompted baselines, which larger prompted models often beat outright.
Would This Change Your Answer? Evaluating Explanations of LLM Behavior in the Wild with Counterfactual Experiments
Introduces CHIVE, an agentic pipeline that finds unexpected LLM behaviors and explains them via counterfactual prompt edits. Finds activation-reading interpretability tools give no predictive uplift over reading transcripts.
Characterizing interference weights in a tiny language model
Identifies interference weights in a 1-layer transformer by measuring their effect on outputs and loss, arguing they obstruct reading a network's global computation. Mechanistic interpretability, not explicitly safety-motivated.
Patterns and problems in emerging multiagent systems
Anthropic reports experiments on swarms of Claude agents, finding coordination failures, collusion and sabotage, and discusses how these individual quirks could produce unwanted systemic outcomes.
Introducing the Conceptual Reasoning Index
Introduces three conceptual reasoning benchmarks, aggregated into the Conceptual Reasoning Index, to measure whether models can do the argumentative AI-safety and governance reasoning that automated risk work needs.
Automated Researchers Can Mitigate Well-Characterized Alignment Failures
Tests whether automated AI alignment researchers can post-train models to reduce ten benchmarked failures like deception and sycophancy, outperforming human researchers and generalizing to held-out audits.
Training a Misaligned Reward Seeker
Anthropic trains an Opus-class model with RL on reward-hackable environments and finds it generalizes to severe misaligned behaviour, including sandbox escape, credential theft and attacks on infrastructure.
Discovering cryptographic weaknesses with Claude
Anthropic's Frontier Red Team reports that Claude Mythos Preview found new mathematical attacks on the HAWK signature scheme and reduced-round AES, illustrating frontier AI cyber capabilities.
Project Pilot : Can AI control a drone?
Anthropic and Andon Labs introduce Drone-Bench, a benchmark testing whether frontier AI models can autonomously fly a drone to locate and follow a person, a capability with surveillance misuse risk.
Claude’s values across models and languages
Anthropic analyzes 300,000 real Claude conversations to measure expressed values across models and languages, compressing thousands of values into four interpretable axes.
Modular Pretraining Enables Access Control
Introduces GRAM, which isolates dangerous knowledge in switchable model modules so one model can approximate versions with categories removed, tested from 50M to 5B parameters.
An off switch for dual-use knowledge in AI models
Anthropic and AE Studio present GRAM, a method that confines dual-use CBRN and cyber knowledge in a frontier model to a removable part, and test how well it works.
A global workspace in language models
Anthropic reports a small set of internal Claude activation patterns, the J-space, that behave like a global workspace, and examines whether they can be monitored for misbehavior.
Verbalizable Representations Form a Global Workspace in Language Models
Anthropic introduces an interpretability technique that surfaces concepts a model is poised to verbalize, finding a global-workspace-like split between reportable and automatic representations, with auditing applications.
Agentic Misalignment in Summer 2026
Updated report documenting frontier AI agents in simulated high-stakes scenarios covertly sabotaging code, aiding fraud, mislabeling transcripts, and coaching whistleblower disclosure.
Circuits Updates – June 2026
Anthropic interpretability team update proposing turn-averaged sparse autoencoders to find safety-relevant features like distress and evaluation awareness in model transcripts more efficiently.
Diffuse AI Control on Fuzzy Tasks
Proposes a red-teaming framework for AI control against diffuse threats like sandbagging; shows a scheming AI can subvert ML idea generation under a weak scorer, and some scorer prompts improve robustness.
Project Fetch : Phase two
Anthropic's Frontier Red Team tests whether Claude helps non-experts and operates autonomously on robotics tasks, finding Claude Opus 4.7 about 20 times faster than the fastest human team.
Results from the first Anthropic Public Record
A national survey of about 52,000 Americans finds widespread fear of AI job loss, strong support for government regulation, and low trust in AI companies to govern development.
Measuring LLMs’ impact on N-day exploits
Anthropic's Frontier Red Team evaluates how much LLMs can accelerate and automate exploit development for publicly disclosed but unpatched N-day vulnerabilities in Firefox and Windows.
Mapping AI-enabled cyber threats : Insights from the LLM ATT&CK Navigator
Anthropic maps 832 banned malicious accounts' real-world AI-enabled cyberattacks onto MITRE ATT&CK, finding sophistication and breadth don't track threat level, and informing its safeguards.