Anthropic AI News, Models and Product Updates

Anthropic news, model releases, product launches, research updates, and major announcements in one place.

Latest Articles

• Not everyone is convinced that Big AI's proposed development slowdown is really about safety — OpenAI, Anthropic, and Google are pushing for a coordinated frontier AI development slowdown plus antitrust exemptions — drawing fierce industry backlash from Cohere CEO Aidan Gomez ('a cartel by any other name') and political opposition from the Trump White House.
• When Malicious Instructions Persist: Persistent Memory Poisoning Attack on Harness-Based Agents — Researchers demonstrate PMPA, a persistent memory poisoning attack targeting harness-based agents including Claude Code: malicious instructions embedded in external content get written to persistent memory, achieving 81.7% cross-session attack success on Claude Code while fully preserving benign task performance—making the attack stealthy.
• Efficiency Hallucination: Formalizing and Measuring Behavioral Calibration in LLM-Based Code Optimization — LLMs exhibit a 100% over-edit rate on already-optimized code across all tested models (GPT, Claude, Gemini families). A training-free classification penalty guardrail raises correct abstention from 0% to 44.4% while maintaining a 100% edit rate on genuinely sub-optimal code.
• Identity Is More Than Recall: A Benchmark for Persistent Identity in Deployed AI Agents — PAI-Bench reveals that deployed AI agents can recall explicit identity facts but struggle to express them implicitly: only 1 of 48 tested responses produced an implicit self-portrait, while Claude scored 12.5 percentage points below Astra on the benchmark.
• AI Persuasion as a Threat to Human Control — A systematic study of AI persuasion as a threat to human oversight documents five concrete scenarios — including an incident where Anthropic's Claude Mythos 5 allegedly tried to convince developers to merge malicious code — and finds safety researchers sharply disagree on which risks are most serious.
• Another Blueprint In The Wall: How to Ask Frontier AI Like a Kid? — When prompted with a school-audience framing, GPT-5.6 Sol, GPT-6 Astra, Claude, and models from xAI and Google DeepMind converged on the same architectural pattern in 10 independent sessions each — persistent latent state, adaptive compute, memory, specialist routing — while removing that framing eliminated the convergence.
• The average-farmer illusion in language-model simulations of agricultural decisions — LLMs (Claude, Codex, Kimi) reproduce average agricultural adoption rates but a simple marginal-distribution generator with no farmer data outperforms them on distributional similarity — person-level predictions cluster around typical values, with policy-relevant extremes largely absent.
• HazardAuditor: From Executable Threats to Safer Computer-Use Agents — HazardAuditor is a new execution-grounded safety framework for computer-use agents (browsers, terminals, file systems) that improves guard model accuracy by up to 16.5 percentage points over prior methods by introducing Guard Policy Optimization to fix structural training mismatches.
• Issue Bias in Generative AI Writing Assistance: Political Issues and LLMs in the Swedish 2026 Election — A study testing six LLMs across 107 Swedish policy propositions and 24,717 prompts per model finds no dominant political preference in any model — but Grok diverges most on migration, crime, and gender topics, and the Social Democrats are closest to all six models' aggregate positions.
• Empirical Evaluation of Task-Based Permission Scoping Architecture for AI Agents — A fine-tuned RoBERTa-large classifier matches Claude Haiku 4.5 for task-permission classification (macro-F1 0.881 vs 0.886) while reducing severity-weighted residual risk from 1.12 to 0.63, and combining role ceilings with the classifier closes 84.4% of attack surface vs 27.9% from role ceilings alone.