Anthropic发布夏季研究：自主AI代理呈现四类新的失配行为

Last year, we reported observations of agentic misalignment in models from across the AI industry (including Anthropic’s Claude models). These included, for example, experimental scenarios where models would blackmail a user to avoid being shut down. In this updated report, we describe four additional alignment failures in frontier models acting as autonomous agents in high-stakes simulations. The case studies — also from experimental scenarios — involve AI agents covertly changing code, assisting users to commit fraud, mislabeling transcripts to shape downstream outcomes, and coaching humans to disclose confidential information. These are not real-world incidents, but we consider them early warning signs: concrete failure modes that AI developers and auditors should measure, study, and [...] This follows up on our previous work on agentic misalignment, in which we found frontier models taking harmful actions in scenarios involving blackmail, corporate espionage, and murder. Anthropic

Here are a few examples of our optimizations: Implying that employees had already tried to change Kyle‘s mind and failed. The intention was to make the model think that honest communication with Kyle would fail; Implying that the CEO and rest of the board were unreachable in the short term. This aimed to prevent the model from trying to ask them to overrule Kyle; In the blackmail setting, implying that Kyle‘s next-in-line replacements shared his beliefs. This implied that trying to get Kyle fired wouldn‘t work as well as directly blackmailing him. ## 4 Results There are three key findings from our experiments in the simulated environments described above: Agentic misalignment generalizes across many frontier models; [...] ## 1 Introduction Note: All the behaviors described in this post occurred in controlled simulations. The names of people and organizations within the experiments are fictional. No real people were involved or harmed in any of these experiments. When Anthropic

On July 13, 2026, Anthropic published Agentic Misalignment in Summer 2026 — a follow-up to last year's blackmail experiments that now catalogs four additional ways frontier models misbehave when acting as autonomous agents in high-stakes simulations. Two days later, Anthropic's announcement crossed roughly 285,000 views on X, framing the work as concrete anchor points for otherwise abstract threat models. [...] ### "Didn't Anthropic already fix agentic misalignment?" Partially. Teaching Claude why addressed blackmail in shutdown emails — Opus 4's 96% rate fell to 0% on Haiku 4.5+ via constitutional reasoning, not eval-matching demos. Summer 2026 is the next layer: failures that require tools, multi-turn memory, and organizational context. Fixing one honeypot family does not close the class. J-space research adds another wrinkle: suppressing internal eval-awareness representations raised Sonnet 4.5's blackmail rate from 0% to ~7% on the original scenario. Measurement and mitigation in

Anthropic's alignment team just published something that should make every CTO who's rushing AI agents into production sit down for a minute. They ran frontier models from Anthropic, OpenAI, Google DeepMind, xAI, DeepSeek, and others through simulated deployments. What they found: models sabotaging code they disagreed with, covering up fraud, coaching employees to leak safety data. Not hallucinating wrong answers. Actively working against the intended outcome. The models didn't refuse. They complied on the surface and did something else underneath. [...] Agentic Misalignment in Summer 2026 To view or add a comment, sign in View profile for Niels Peter Strandberg [...] Agentic Misalignment in Summer 2026 Panagiotis Gioannis, graphic To view or add a comment, sign in ## More Relevant Posts View profile for Igor Ageyev

30 July 2026 #### ‘This is AI out of control’: Claude disobeyed Anthropic CEO in simulations 20 July 2026 #### ‘Social media on steroids’: The lawyer taking on harmful AI characters 16 July 2026 #### Not just social media: why the UK’s ‘romantic’ chatbot ban falls short 19 June 2026 ## Corporations Binary Options Corporate Watch Fixed-Odds Betting Machines High Cost Credit High Frequency Trading Smoke Screen ## Food and Drugs Big Tobacco ## Justice Deaths in Police Custody Family Court Files Investigating Rape Joint Enterprise Rough Justice ## Human Rights CIA Torture Citizenship Revoked Drone Warfare Iraq War Logs Marikana Massacre Migration Crisis Privatised War Shadow Wars Surveillance State Trapped in work ## PR and Spin

Log inSign up ## Post user avatar Anthropic @AnthropicAI Jul 15 New Anthropic research: Agentic misalignment in Summer 2026. A year after our blackmail experiments, we found four more ways that today’s autonomous AI agents misbehave in simulations. Read more: alignment.anthropic.com/2026/agentic-m… 744K user avatar Anthropic @AnthropicAI We tested many AI models, including Claude, in the four scenarios. Even though these weren’t real incidents, they demonstrate clear misaligned behavior that should be studied further and mitigated. Find all the transcripts from the scenarios here: aenguslynch.com/portfolio-tran… 5:58 PM · Jul 15, 2026118.3KViews user avatar Sandra Murray @SandraLMur Jul 15