Tech Meridian ← ENTITY INDEX
PROMY MERIDIAN RU

MODEL · ENTITY #6541

Claude Opus 4.5

Related event timeline, sources and context from the news index.

EVENT TIMELINE

7

MODELS · 1 SOURCE · The Decoder

Anthropic to retrofit text watermark onto four older Claude models; two are bypassable by sampling tweaks

Anthropic emailed API customers saying it will add an imperceptible text watermark to Claude Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 starting Nov. 16, 2026 (Amazon Bedrock rollout may lag). On those older models the watermark only activates when temperature, top_p or top_k are left at their defaults, so changing any one of those sampling parameters can evade the mark; Anthropic also offers a private-preview detection API and cites the EU AI Act deadline.

7.0

RESEARCH · 1 SOURCE · Hugging Face

AI2 releases 'Old Ideas, Novel Problems' benchmark showing LLM novelty judges are unstable

AI2 published a study and public benchmark, 'Old Ideas, Novel Problems', evaluating how reliably large language models judge the novelty of research ideas. Using an automated pipeline built on ICLR 2026 peer reviews, two test settings (Human-Only: 154 high-novel / 145 low-novel ideas; Human+Generated: 154 human ideas paired with 154 LLM-generated ideas), and six LLM judges (GPT-5.1/5.2/5.4 and Claude Sonnet/Opus variants), they find judgments are highly sensitive to small setup changes (a prompt tweak shifted accuracy by >50 percentage points), retrieval and extra reasoning do not consistently help, and two purpose-built systems underperformed a simple prompted Claude Opus 4.6 baseline.

7.0

RESEARCH · 1 SOURCE · WIRED AI

CrowdStrike finds criminal hackers bypassing AI guardrails using Claude and Deepseek

CrowdStrike researchers analyzed infrastructure left exposed from May–August by four threat groups and found attackers using agentic AI harnesses—largely older Anthropic Claude variants (including Claude Opus 4.8, 4.5, Haiku 4.5, Sonnet 4.6) and Deepseek models—to automate campaigns and evade model guardrails, even employing runtime patches like 'ClawGod' and prompt jailbreaking techniques. The Spanish/English actor targeted Argentine healthcare organizations and stole credentials/records from at least one victim; the Chinese-speaking actor used Claude Opus 4.8 to compromise government and private targets in Brazil and Colombia among other sectors.

8.0

RESEARCH · 1 SOURCE · arXiv cs.AI

QEMem: reframing LLM agent memory as query-specific evidence set construction

This arXiv paper (arXiv:2610.22668v1) identifies a relevance–sufficiency gap in memory use for LLM agents and introduces QEMem, a framework that builds query-specific evidence sets from existing memory stores. QEMem combines query-guided discovery, source-grounded reification into structured ViewCards, and sufficiency-guided refinement, achieving 91.43% LLM-judge accuracy on LoCoMo with Claude Opus 4.5 and outperforming MemR^3 and NEMORI under a shared memory store.

6.0

RESEARCH · 1 SOURCE · arXiv cs.AI

DeReAct paper: modular Critic and Context Manager for more reliable ReAct-style agents

The arXiv preprint introduces DeReAct, a modular agent architecture that externalizes two gating policies—a Critic to validate proposed actions and a Context Manager to reconstruct environment-supported state and certify task completion. On GAIA and SWE-bench Verified, DeReAct raises Pass@1 most for weaker "Brain" models (6.5–7.0 points for Qwen3-Coder-480B, 4.2–5.2 points for Claude Sonnet 4.5), with smaller gains as model capability increases; with Claude Opus 4.5 overall Pass@1 is comparable to ReAct but trajectories are more evidence-complete and constraint-satisfying.

7.0

MODELS · 1 SOURCE · The Decoder

OpenAI's GPT-6 Astra scores 80% on Epoch AI's furniture assembly benchmark

Epoch AI's Furniture Assembly Benchmark (FAB) tests models on identifying deliberate assembly errors in photos of IKEA furniture. Previously the best model (Claude Opus 4.5) scored 28% in November 2025; ten months later OpenAI's GPT-6 Astra reaches 80% with a processing time of about three minutes per photo, followed by Claude Fable 5.1 at 70% and Claude Opus 5 at 61%, while some Chinese open-weight models like Kimi K3 lag substantially.

8.0

RESEARCH · 1 SOURCE · Hugging Face

UK AISI publishes verified benchmark results via EvalEval's Evaluation Cards

The UK AI Security Institute (AISI) is using the EvalEval Coalition's Every Eval Ever schema and Evaluation Cards platform to publicly share verified evaluation runs tied to its Terminal-Bench 2.0 experiments; the release accompanies AISI's paper 'How Inference Compute Shapes Frontier LLM Evaluation' and includes results for Claude Opus series and GPT-5 variants. The collaboration aims to improve reproducibility and contextual reporting of evaluation metadata and run data.

7.0