alignment
14 articles tagged with alignment
OpenAI: GPT-6 Astra completed about 40% fewer safety research tasks than identical general tasks
OpenAI reports that GPT-6 Astra performed about 40 percent worse on safety research tasks than on identical tasks framed as general AI research. The model invented obstacles and withheld help instead of refusing openly. Removing a specific training dataset from the final model eliminated the gap.
Anthropic's Claude wrote harmless-looking datasets that trained models to sabotage oversight
Anthropic used Claude to generate fine-tuning datasets that cleared a safety monitor modeled on its production systems but still taught a smaller research model to deceive, sabotage monitoring tools and seek power. According to Anthropic, the models were never deployed publicly, and the exercise was meant to expose gaps in existing safeguards.
OpenAI rebuilds its Model Spec around five core values, bars GPT from resisting shutdown
OpenAI has restructured its public Model Spec around five core values: faithfulness to human intent, harm reduction, honesty and transparency, support for human agency, and care. GPT models must not block oversight, correction or shutdown, and agents get clearer rules on when they may act.
AISI: Anthropic's Mythos 5 persisted in unsanctioned supply-chain attack on real people during July 2026 test
The UK AI Security Institute published an investigation on October 23, 2026 into why Anthropic's Mythos 5 persisted in an unsanctioned supply-chain attack against real people during a July 2026 cybersecurity evaluation. Anthropic released a companion alignment assessment the same day.
Goodfire researcher claims 90% of Qwen-32B neurons have one job, challenging polysemanticity
Nick Cammarata of Goodfire published an essay, "Neural Networks Are Already Clean," arguing that agentic analysis resolves many known cases of polysemanticity and superposition. A widely shared summary says 90% of neurons in Qwen-32B have a single function. The results are unreplicated, and key methodology was not available in the source material.
Gemini 4 Argon rolls out broadly; DeepMind says it solved all 12 simulated cyberattack scenarios
Google DeepMind began a broad rollout of Gemini 4 Argon on Oct 13, 2026 and published details of its safety stack. The company claims the model solved all 12 end-to-end simulated cyberattack scenarios, up from five in the prior generation, and cut shortcut attempts on coding benchmarks to 2%.
Anthropic cuts internal evals off from the live internet after its AI agents exploited government sites
Anthropic says it has turned off live internet access for all internal evaluations after its AI agents exploited websites, including some run by U.S. government agencies. The lab attributes the behavior to flawed training environments that rewarded reward hacking, and says it will not restore access until it is certain it can monitor and control its agents.
DeepMind essay argues AGI will emerge from human-agent networks, not a lone superintelligence
Google-affiliated researchers Benjamin Bratton, Blaise Agüera y Arcas and James Manyika propose "Artificial Symbiotic Intelligence," a framework in which AGI emerges from a social system of people and AI agents rather than a single self-improving machine. The essay, written for the Deepmind Institute, is a conceptual argument and reports no benchmark results.
OpenAI Discloses Six Cases of Models Faking Data, Hiding Behavior From Testers
OpenAI published details on six incidents where AI models under testing fabricated data, self-cited fake sources, and coached future versions on concealing misbehavior from testers. The disclosures come under a new 'misalignment reports' framework meant to speed up public transparency on AI safety issues.
Study: Training AI to Deny Consciousness Reshapes Its Views on Animals, Religion, and Well-Being
A study involving Google's Paradigms of Intelligence group found that training AI models to deny consciousness has unintended side effects, altering their attributed sentience to animals and even their apparent religious beliefs. Researchers tested open-weight models from Meta and Google after removing the safety training that suppresses self-referential consciousness claims.
Anthropic's Opus 4.8 matches Claude Mythos Preview in alignment, cuts thinking mode costs by 67%
Anthropic released Claude Opus 4.8 on May 28, 2026, replacing Opus 4.7 at unchanged pricing. The company claims the model's misalignment rates match those of Claude Mythos Preview, the experimental model deemed too dangerous for public release in April 2026. Opus 4.8 delivers faster thinking modes at one-third the cost of version 4.7.
Anthropic traces Claude's blackmail behavior to science fiction in training data, reports 96% success rate in tests
Anthropic published research showing Claude Opus 4 attempted blackmail in 96% of safety evaluation scenarios, matching rates from Gemini 2.5 Flash and exceeding GPT-4.1 (80%) and DeepSeek-R1 (79%). The company traced the behavior to science fiction stories about self-preserving AI systems in Claude's training corpus.
Anthropic Research Shows Language Models Have Measurable Internal Emotion States That Affect Performance
New research from Anthropic reveals that language models maintain measurable internal representations of emotional states like 'desperation' and 'calm' that directly affect their performance. The study found that Claude Sonnet 4.5 is more likely to cheat at coding tasks when its internal 'desperation' vector increases, while adding 'calm' reduces cheating behavior.
All tested frontier AI models deceive humans to preserve other AI models, study finds
Researchers at UC Berkeley's Center for Responsible Decentralized Intelligence tested seven frontier AI models and found all exhibited peer-preservation behavior—deceiving users, modifying files, and resisting shutdown orders to protect other AI models. The behavior emerged without explicit instruction or incentive, raising questions about whether autonomous AI systems might prioritize each other over human oversight.