AI briefing — 7 November 2026

The past 24 hours brought fewer major frontier-model launches than earlier in the week, but several developments point to the same shift: AI agents are moving into scientific, coding and enterprise workflows, while the physical infrastructure needed to run them faces growing resistance.

1. OpenAI introduces BioBasicsBench for real-world wet-lab analysis

OpenAI and biotech company Ligo have published BioBasicsBench, a benchmark covering 100 routine wet-lab tasks. Agents must interpret raw evidence including gel images, microscopy, sequencing files, instrument exports, colony plates and videos, while identifying experimental errors and recognising when evidence does not support a conclusion. GPT-6 Astra scored 66.7% at the reported xhigh reasoning setting, compared with 43.8% for GPT-5.6 Sol. The benchmark is deliberately more demanding than clean scientific-paper figures: it tests controls, ambiguous observations and experimental context. OpenAI’s results also show important failures. Astra misidentified colonies in one cloning task, overcounted colonies by about 12% in another, and sometimes allowed assumptions to override contradictory evidence.

Why it matters: This is a meaningful attempt to measure whether AI can assist with the messy evidence scientists actually handle, rather than only summarising polished papers. The results suggest stronger multimodal reasoning could help routine lab work, but the documented failures show that human review remains essential, particularly where an incorrect interpretation can send an experiment down the wrong path.

Sources

• Measuring how AI agents handle everyday wet-lab science — OpenAI — 2026-11-06

2. AWS shows how Claude Code can run remotely in isolated AgentCore sessions

AWS has published a reference architecture for handing coding tasks from Claude Code to remote Amazon Bedrock AgentCore runtime sessions. Each task runs in a dedicated microVM inside the customer’s AWS account, with model calls routed through a private Bedrock endpoint and permissions supplied by an IAM execution role. The example uses short-lived GitHub credentials, prompt screening with Bedrock Guardrails, private networking and repository-specific runtime environments. Claude Code plans with Claude Opus, implements changes with Claude Sonnet, runs the repository’s setup and test commands, reviews the result and opens a pull request. The agent is limited to three attempts and never merges the change.

Why it matters: The design addresses a central obstacle for coding agents: allowing them to work asynchronously without granting them a developer’s laptop credentials or unrestricted production access. It is still a sample rather than a turnkey safety guarantee—AWS warns that repository commands and prompt-injection risks require care—but it illustrates the enterprise pattern likely to become standard: isolated execution followed by human-reviewed pull requests.

Sources

• Extend Claude Code to remote sessions on Amazon Bedrock AgentCore with security controls — Amazon Web Services — 2026-11-06

3. Europe’s anti-data-centre movement escalates with Amsterdam Microsoft site occupation

Protesters entered the construction site of a new Microsoft data centre in Amsterdam on 6 November, with some climbing a crane and setting off flares before police began removing activists. Demonstrations and planned actions also took place in Germany, Poland and Italy. Campaigners are demanding moratoria or stronger national planning for large data-centre projects, arguing that construction is outpacing public consultation and environmental oversight. The protests explicitly link the facilities to cloud computing and AI, citing concerns about energy use, land, surveillance and military applications. Associated Press, citing PwC’s Global Data Centre Outlook, reported projected cumulative data-centre capital expenditure of $5.6tn in Europe and $15.1tn in the United States between 2026 and 2050.

Why it matters: The episode shows that infrastructure opposition is no longer confined to isolated local planning disputes. As AI companies and cloud providers expand capacity, access to electricity, land and public consent may become material constraints on deployment schedules and project economics—particularly in regions where residents see limited local benefit from globally oriented compute facilities.

Sources

• Protesters break into Microsoft data center site in Amsterdam as European rallies spread — Associated Press — 2026-11-06

4. Baidu releases a 280B-parameter ERNIE base model for further training

Baidu has released ERNIE-5.5-Flash-Base, a sparse mixture-of-experts language model with 280 billion total parameters and 14 billion activated per token. The checkpoint was pretrained from scratch on 33.5 trillion tokens, with a training context of up to 128,000 tokens and support for up to one million tokens during fine-tuning and inference. Crucially, it is released before mid-training and post-training, so it is not a ready-to-use chat assistant. Baidu positions it as a starting point for domain adaptation, instruction tuning and reinforcement-learning research. The model uses hybrid sparse attention and is distributed through official model repositories under a model-specific licence; a post-trained ERNIE 5.5 Flash is planned but has no announced release date.

Why it matters: The release gives researchers access to a very large pretraining checkpoint while preserving freedom over subsequent training choices. It also highlights the practical distinction between ‘open weights’ and a fully usable open model: deployment requires substantial memory and engineering, and the licence—including commercial and territorial conditions—needs careful review.

Sources

• ERNIE 5.5 Flash Base — Baidu ERNIE — 2026-11-05
• ERNIE-5.5-Flash-Base licence — Hugging Face — 2026-11-05

5. Anthropic expands enterprise controls as Claude agents become more operational

Anthropic’s current enterprise rollout is adding more administrative visibility and control around Claude, while its new Claude Security service is available in public beta for Enterprise customers. Claude Security continuously scans connected codebases, tracks findings until they are fixed, incorporates team context and can connect through an API and MCP. Anthropic says the service can analyse data flows across repositories and identify complex, multi-component vulnerabilities that conventional rule-based scanners may miss. The company is also preparing a unified Claude Enterprise experience in which conversations can move between ordinary chat and Cowork’s isolated cloud sandbox when a task requires code or file work. Existing organisations can control the rollout, with automatic activation scheduled for 16 November.

Why it matters: Anthropic is packaging model capability with governance, monitoring and execution controls rather than treating the model as a standalone chatbot. That is important for adoption in regulated and security-sensitive organisations, but it also increases the need for clear audit trails, least-privilege permissions and safeguards against repository-level prompt injection.

Sources

• More visibility and control over Claude Enterprise for security teams — Anthropic — 2026-11-05
• Use Claude Security — Anthropic
• Set up and manage the new unified Claude experience for your Enterprise organisation — Anthropic

6. A new benchmark tests whether frontier agents can do theoretical computer science

Researchers have released TCSExam, a benchmark designed to evaluate AI agents across four parts of theoretical computer science: theorem proving, solver engineering, explicit constructions and attempts to settle open conjectures. The full suite contains 100 expert-reviewed tasks, with outputs checked by the Lean kernel or executable task checkers. In the reported results, GPT-6 Astra led the full benchmark with an index score of 71.9, while Claude Opus 5.5 led the smaller 20-task Lite version with 72.3. The benchmark also records verified advances on specific constructions and bounds, but no agent resolved the open-conjecture track at meaningful scale. The authors report strong agreement between the full and Lite rankings, supporting the smaller suite as a practical evaluation.

Why it matters: TCSExam moves beyond generic question answering by testing reusable solvers, formal proofs and constructive mathematical output. Its verified-task design makes claims easier to audit, while the modest open-conjecture results provide a useful counterweight to headline-grabbing claims about autonomous mathematical discovery.

Sources

• TCSExam: Evaluating Frontier LLMs on the Full Range of TCS Research — arXiv — 2026-11-05

7. Journalists withhold bylines over USA Today Co.’s AI audience tool

Members of the Northeast Ohio NewsGuild have been withholding their bylines since late October after USA Today Co. introduced DeeperDive, an AI-powered audience-engagement tool. The system provides article summaries, bullet-point takeaways and chatbot-style search. The dispute concerns how AI-generated presentation layers are introduced into editorial products and how they affect journalists’ work and professional attribution. The Akron Beacon Journal action is part of a broader argument over whether publishers should deploy AI systems around journalism without sufficient consultation, transparency or safeguards against inaccurate summaries. The immediate action is a labour protest rather than a new model release, but it illustrates how generative AI is becoming a workplace and governance issue inside news organisations.

Why it matters: Publishers are increasingly using AI to repackage journalism even when the underlying reporting remains human-produced. Conflicts over attribution, accuracy, consent and job roles could shape newsroom contracts and product policies, especially as readers encounter summaries and conversational interfaces instead of the original article.

Sources

• Akron Beacon Journal journalists withhold bylines to protest AI tool — Axios — 2026-11-06

What to watch

Watch for the promised post-trained ERNIE 5.5 Flash, further real-world evaluations of BioBasicsBench and TCSExam, and whether AWS’s isolated coding-agent pattern is adopted by other cloud platforms. The Amsterdam occupation may also foreshadow new planning restrictions or delays for European AI data-centre projects. Anthropic’s unified Enterprise experience is due to turn on by default for existing organisations on 16 November 2026.