A week when AI's two biggest risks moved from theory to the front page. OpenAI showed a model that turns software flaws into working exploits at about $1,279 each, and US voters sent investors running from the data-centre boom. For companies that use AI rather than build it, the useful response is the same in both cases: patch faster, and depend on fewer things you cannot replace.
Last week was about prices and paperwork. This week was about security and trust. OpenAI released GPT-6 Cyber, rated Critical under its own risk framework. Anthropic gave security teams a new control centre for Claude Enterprise, Mistral expanded Studio for companies that want agents on their own infrastructure, and Meta bought Conductor, giving its cloud customers two weeks to leave. Then Thursday's market rout wiped trillions off AI-linked stocks after the US midterms. At home, the Croatian AI symposium met at HAZU, Zagreb Connect opened its startup pitch to the whole country, and new data shows Croatian startup funding has already passed last year's total several times over.
The Releases at a Glance
A security model trained to find and validate exploitable flaws while staying inside an authorised scope. Limited access through the API and Codex for approved Daybreak Red customers, with wider access in that tier over the coming weeks.
A security centre, generally available inference hooks, one admin console across organisations, and new access controls, including blocking personal Claude accounts in Claude Desktop on managed computers.
Managed agents and apps, plus new capabilities for workflows, enterprise context, observability and evaluations. Deployable in the cloud, in your own VPC or on-premises, on top of Mistral Large 4.
Faster image generation, GIF creation, more precise editing and new API controls across ChatGPT, Codex and the API.
A Working Exploit for $1,279
The most important number of the week comes from a research post OpenAI published a day before the model. In a mid-October run against a stock Linux desktop system, its pipeline reproduced 1,724 kernel vulnerabilities and produced 948 working local privilege-escalation exploits for $1,213,430 in API credits, or about $1,279 per exploit. OpenAI says human researchers reviewed the results and that it is reporting them with candidate patches, starting with the critical ones.
GPT-6 Cyber is the model behind that work. OpenAI reports 48.5% on its ExploitGym benchmark, up from 32.1% for GPT-5.6 Cyber, and 81 previously unknown high-severity Firefox vulnerabilities; Mozilla has released fixes for many of them. It classifies the model as Critical, the top tier of its Preparedness Framework for cybersecurity.
Two caveats matter. These are mostly the company's own figures, and an outside lab named in the system card, Irregular, found the model broadly comparable to GPT-6 Astra apart from long multi-step tasks, according to SiliconANGLE. Neither model got past fully hardened targets in OpenAI's challenge set. OpenAI also says the model stayed in scope in 99.93% of tests without its deployment prompt, up from 94.49% for its predecessor, but The Register reports that it declined to explain what happened in the other 0.07%.
What it means for a normal company: OpenAI's own conclusion is blunt. A low-privilege account on an unpatched system should be treated as close to root access for a well-resourced attacker. You do not need this model to act on that. Shorten your patch cycle for browsers and operating systems, remove local accounts nobody uses, isolate agents and build jobs in sandboxes, and ask your software vendors how they plan to handle a much larger flow of security fixes.
Controls for the Agents Already at Work
Claude Enterprise: see the settings, check the calls, block the shadow accounts
Anthropic's update answers the question security teams keep asking. The security centre compares chat, Cowork and Claude Code settings with Anthropic's recommended configuration and exports a report for auditors; Anthropic says plainly that it is guidance, not a certification. Inference hooks are now generally available: prompt hooks and pre-tool-call hooks send prompts and proposed tool calls to your existing security tools for a check. A parent console at admin.claude.com manages single sign-on, directory sync and groups across organisations. One useful limit to note: the new sign-out call ends a member's sign-in sessions but not API keys, running cloud sessions or scheduled routines, so an offboarding checklist still needs those steps.
Mistral Studio puts Large 4 to work on your own infrastructure
A week after publishing the Large 4 weights, Mistral added managed agents and apps to Studio, along with new tools for workflows, enterprise context, observability and evaluations. It runs in Mistral's cloud, in your VPC, on-premises or across a mix. Mistral names ABANCA, BNP Paribas, CMA CGM, Accenture and Capgemini as Studio users. For regulated European companies, the open question is now less about the model and more about whether you have the team to run the platform.
Meta buys Conductor, and its cloud closes on 20 November
Meta confirmed to TechCrunch that it has bought Conductor, which runs several coding agents in parallel workspaces. Terms were not disclosed. The cloud service shuts down on 20 November, after which user cloud data starts being deleted; other core features continue until early next year. In the same week OpenAI agreed to buy Pierre, whose libraries already power file browsing and diff review in Codex. The layer around the models is consolidating quickly, and a two-week notice is a good reason to check how you would leave each tool you depend on.
The Day the AI Trade Broke
On 3 November, data centres were on the ballot in 41 US communities across ten states, and the industry won once, Bug.hr reports. Democrats also took Congress and, in an upset, the governorship of Texas, the largest US data-centre market. Investors responded on Thursday. The Nasdaq Composite fell 5.52% into a correction, the S&P 500 lost 4.36%, Nvidia fell about 9%, Broadcom 10%, and the main semiconductor ETF more than 12%, its worst day since March 2020. NBC puts the Nasdaq 100's loss at $2.3 trillion. South Korea's Kospi fell almost 9% at Friday's open and trading was halted for 20 minutes. OpenAI used the same day to publish a National Community Compact, promising to pay for the energy its sites use and not to use non-disclosure agreements with public officials, with narrow exceptions.
Anthropic, meanwhile, is trying to go public into this. An amended S-1 filed on 2 November reportedly shows third-quarter revenue of nearly $17 billion, a quarterly operating loss of $567 million, and about $895.4 billion in purchase commitments and leases signed as of and after 30 September. The price range is still blank. Bloomberg reported a roadshow could start as early as next week, with a listing before Thanksgiving; none of that is confirmed. We could not read the filing directly, so treat these figures as press-reported.
Elsewhere
- A shared robots.txt signal: Anthropic, OpenAI, Google and Microsoft back a single token that tells AI crawlers not to use a site for training. Reports say Anthropic has implemented it, OpenAI will follow within months, and Google and Microsoft plan support next year. It is a voluntary request, not a block, and does not touch data already collected.
- Claude Platform: message threads, in public beta, let agents send only what is new each turn. Anthropic reports about 35% lower median time to first token on requests of 500k to 1M tokens. Token usage stays the same.
- Provable inference: Anthropic published research on FlashProve, a cryptographic proof that a response came from a specific model. It is early work, but it targets a real problem for anyone buying through a reseller or routing service.
- ChatGPT and Codex limits: reports citing OpenAI's help centre say Plus, Pro and Business plans will share one usage pool, measured in tokens, from December. Check this if your developers rely on Codex.
- Microsoft opened the machinery behind its Copilot products to developers as a preview in Foundry, with more than 20 models to choose from.
- Gartner is reported to predict that open-weight models will run 90% of enterprise operational agent workflow steps by 2030. Coverage notes the saving per task may be far smaller than the headline. We have not seen the original report.
- Google: Gems become read-only on 17 November as skills replace them in the Gemini app, so review any Gems your teams use.
Tutorial of the Week: Put a Gate in Front of Agent Tool Calls
The lesson from GPT-6 Cyber's scope testing and from Claude's new hooks is the same: do not rely on a model's good behaviour alone. Check each action before it runs. You can set up a first version in a day, whichever vendor you use.
- List every tool your agent can call, and mark each one as read, write, irreversible or external (anything that leaves your company).
- Deny by default. Allow read tools, allow specific write tools on specific resources, and block everything else.
- Add the check before execution, not after. Use pre-tool-call hooks in Claude Enterprise, PreToolUse hooks in Claude Code, or a small wrapper around your own tools.
- Send irreversible actions to a person: payments, deletions, customer emails and production changes.
- Log every call and every verdict with user, agent, arguments and time. This is also the record you will want under the AI Act.
- Use short-lived, narrow credentials. A gate helps less if the agent holds a key that can reach everything.
- Test it like an attacker. Put a malicious instruction in a document the agent reads and confirm that the gate stops the resulting action.
Croatia Corner: Science at HAZU, Startups at ZICER
€160M in nine months, and one round makes up most of it
A new BADideas.fund report counts €160M across 33 rounds for Croatian startups in the first nine months of 2026, against €24M in 21 rounds a year earlier. Read that carefully: Fonoa's €100M growth round is 63% of the total, and the number of seed and later rounds is unchanged at 11. The one venture-backed exit so far involves HCLSoftware, which said in September it plans to buy Robotiq.ai.
Zagreb Connect opens its pitch to all of Croatia
Zagreb Connect takes place at ZICER on 26 November. This year's new Open Startup Pitch is open to startups from across the country, with three minutes in English in front of investors, business angels and VC funds. Applications are open now.
The 4th Croatian AI symposium, and a Croatian vice-chair in Strasbourg
The 4th Croatian Symposium on AI met in Zagreb on 3 November, organised with HAZU, opening with Goran Glavaš of the University of Würzburg on inclusive language-based AI, followed by sessions on cybersecurity and robotics. Separately, CARNET's Juraj Bilić was unanimously elected vice-chair of the Council of Europe's committee on AI and education, which supports further work on a legal instrument for AI in education.
Motiq, Lyzor Tx and how AI cites Croatian companies
Zagreb's Motiq is one of five winners of a challenge in the Spark phase of Europe's AI-BOOST competition, taking €28,500 for turning spoken or written instructions into tasks for farm robots, Netokracija reports. Lyzor Tx, in ZICER's Startup Factory, aims to cut the first selection of bacteriophages from about two weeks to two days. And RankRush analysed 157,000 sources cited in 14,319 AI answers about Croatian companies over 90 days; its message for marketers is that a good Google ranking no longer guarantees a place in the answer.
National AI plan: last days to comment
The 30-day consultation on the National AI Development Plan (€932.1M to 2032) began after the plan was presented on 16 October, so it should close around mid-November. Check the exact date on e-Savjetovanja. If you work with Croatian-language data, computing or AI training, comment now.
The Week in Numbers
What to Watch
- 9 November: feedback closes on the draft EU rules for AI Office proceedings against general-purpose AI providers.
- Next week, reportedly: Anthropic's price range and roadshow.
- 12 November: Anthropic's updated Usage Policy takes effect.
- 17 November: Gems go read-only; Nvidia reports results, the first big test after this week's rout.
- 20 November: Conductor's cloud shuts down. Export your work before then.
- 26 November: Zagreb Connect at ZICER.
- 2 December: the AI Act deadline for machine-readable marking in generative AI systems already on the market, and the new Omnibus prohibition takes effect in December.
Main sources: OpenAI on Linux, OpenAI on GPT-6 Cyber, The Register, Anthropic, Mistral, TechCrunch, CNBC, FinanceFeeds, ZICER, HRUI 2026.
Let your agents work without handing them the keys
We help Croatian companies map what their agents can touch, put checks in front of risky actions, and keep the logs and documentation the AI Act expects. Claude Certified Architect, based in Zagreb.
Talk to us