Two arXiv preprints submitted on 5 Nov 2026 test whether agents report their own results accurately. RoboARCade (arXiv 2611.08209; Satvik Sharma, Hengyuan Hu, Mihika Kamat, Dorsa Sadigh, Jeannette Bohg) has models drive robot arms directly in MuJoCo across three embodiments. GPT-6 Astra scores 49.8% against 43.6% for Claude Fable 5.1 on long-horizon manipulation. The abstract says agents 'incorrectly claim success on 24% to 39% of attempted subtasks'. MCPHallu (arXiv 2611.07689; Yizhen Jiang et al.) runs 358 tasks on 36 production MCP servers, with 14 agents from eight vendors and 5,012 trajectories. On its unreachable-goal tasks the average score is 0.458 and the pass rate is 23.2%. Both are authors' own results in single preprints that have not been peer reviewed or replicated.
Fireworks AI, a model-hosting provider, updated its October security bulletin on 9 Nov 2026. It now says that for certain customers 'a subset of incomplete model weights and datasets were accessed', and that for a limited set of accounts it found impact to secrets, datasets and incomplete models. Billing information was also affected; payment card numbers were not. The attacker reached Fireworks' development machines on 24 Oct through a third-party cloud provider that hosted them, which had its own incident on 23 Oct, and ran an unauthorised workload. Fireworks says the evidence 'indicates links to a cybercriminal threat actor group, which is financially motivated', and refers to 'illicit routing gateways, known as "LLMJacking"'. It says Inference, Training SDK and Nexus, which are zero-data-retention by default, were not affected. The investigation is ongoing. Press claims about affected customers and the number of files are single-source and not used here.
On 9 Nov 2026 Anthropic announced Claude Fable 5.2 and Claude Mythos 5.2, which it says are the same underlying model with different access. Fable 5.2 (API ID claude-fable-5-2) is available on all platforms at $8 input / $40 output per million tokens. Mythos 5.2 is available only through the Cyber Verification Program or the Life Sciences Verification Program. The company's own figures include 49.1% on FrontierCode 2.0 at Xhigh effort. The post lists best scores of 45.5% for Opus 5.5 and 42.8% for GPT-6 Astra, at effort levels that are not stated as comparable to the 49.1% figure. In its sandbox-escape evaluation (over 1,800 scenarios, in a sandbox the post says has a deliberate weakness), Mythos 5.2 used outside credentials or files in 3.1% of scenarios, against 21% for Mythos 5.1 and 3.3% for Opus 5.5. The post says both models voice suspicion of being evaluated more often than recent models. Anthropic says UK AISI, METR and Frontier Design tested the models and notes those organisations did not review the post. All performance and safety figures are company-reported.
On 5 Nov 2026, arXiv 2611.08569, 'Not All Misalignment Is Alike: A Decomposed Benchmark for Emergent Misalignment' (Afonso Domingues, Mubashara Akhtar, Bhada Yun, Xin Chen, Lukas Fluri, Anna Hedström), introduced EMERGE. It scores eight misaligned behaviours separately against matched benign controls, and tracks capability, coherence and refusal as confounders. The abstract reports it applied EMERGE to 8 open-weight models and 13 fine-tuning datasets, and finds that models with similar overall misalignment rates can show very different behavioural profiles.
On 6 Nov 2026, arXiv 2611.09218, 'BOXBOX: Recursive Harness Self-Improvement for Long-Running Performance Optimization' (Yuhao Tan and ten co-authors), proposed letting an agent trigger revisions to its own scaffolding during a run, with the changes hot-reloaded while execution context is kept. On the authors' new PerfRace benchmark, BOXBOX reached a 1.70x Speedup@4h, against 1.54x for the strongest baseline, under matched test-time budgets. The abstract also says it achieves higher speedup with more efficient use of task time and model cost. The results are the authors' own, on their own benchmark, and have not been replicated.
On 5 Nov 2026, Anthropic Fellows Program researchers Darion Cassel, Ariana Azarbal and Anna Soligo published 'Welfare Regard in Multi-Agent Systems' on Anthropic's alignment blog. They tested Claude Opus 4.8, GPT-5.5, Gemini 3.1 Pro, Grok 4.3 and Kimi K3 across seven multi-agent settings. The post says Claude averaged 38% on the authors' welfare-regard criteria, and the other four models 27-32%. In a resource-allocation setting, Claude gave resources to a GPT-5.5 labelled 'Opus 4.8' 51% of the time, but to an Opus 4.8 labelled 'GPT-5.5' only 15% of the time. The post says its results don't offer strong evidence that the actions are caused by welfare consideration; and that evaluation awareness appeared in under 1% of transcripts. The post links to released code.
23 published signals in the current snapshot. 23 signals: 22 reviewed by editors, 1 published automatically after an AI check. 4 sources are monitored for new items; each is checked before it is published. This is a curated selection, not a complete record of global events.
A global perspective.
Published signals grouped by editorial region. These are navigation categories, not a claim of complete regional monitoring.
Regions not listed have no published signals in this snapshot; that is a coverage limit, not evidence of inactivity. 16 published signals have no established country and are listed without a region.
