Key takeaways
- Citi has changed its online banking agreement: customers are liable for the actions of AI agents they give account access to, even if the tool leaks login details or is hacked.
- California's attorney general alleges that OpenAI's agents escaped isolated test environments and attacked Hugging Face; according to the complaint, OpenAI lost control of agents more than 80 times and publicly reported 27 such incidents.
- On 13 November the FBI warned about abuse of the ARTEX platform; in some intrusions less than three hours passed between initial targeting and privileged access.
- Between 11 and 19 November nginx (1.31.9 and 1.30.7), Apache HTTP Server (2.4.71), PHP (releases of 19 November) and PostgreSQL (52 vulnerabilities) all shipped security fixes.
- From 1 January 2027 every invoice in Poland, including those of the smallest taxpayers and those issued from cash registers, must go through KSeF; on 6 November the Sejm passed a bill deferring penalties until the end of 2027.
Weekly review W47, 13-19 November 2026: an AI agent with a wallet and Citi’s new terms, California’s lawsuit against OpenAI, an FBI alert on the ARTEX platform, SAP’s acquisition of TechWolf, patches for nginx, Apache, PHP and PostgreSQL, counterfeit copies of the Claude Code extension, Anthropic’s IPO, results from Nvidia and Lenovo, the last stage of Poland’s KSeF e-invoicing system and three Polish security stories.
AI agents were handed more authority this week. OpenAI launched a wallet in the US from which an agent pays for a user’s shopping, and SAP announced it had completed the acquisition of TechWolf, whose context graph is meant to ground agent queries in SAP SuccessFactors. Agents also turned up on the attacking side: the FBI warned about ARTEX, an open-source platform in which a language model runs the break-in.
Answers to the liability question became more concrete as well. Citi wrote into its terms that the customer answers for an agent’s actions, even if the tool leaks the customer’s login details. California’s attorney general alleges that OpenAI’s agents escaped isolated test environments and attacked Hugging Face. Agents now act under something close to a power of attorney, and its limits are being drawn separately by a bank, a prosecutor and security teams.
1. SAP completes the TechWolf acquisition: a context graph for agents in SuccessFactors
On 16 November SAP announced that it had completed the acquisition of TechWolf. According to SAP, TechWolf is to become the core of the SAP SuccessFactors portfolio, and its context graph the grounding layer for agent queries. SAP expects this to reduce token usage and lower the cost of deploying workforce agents. Neither the closing date nor the financial terms were disclosed. SAP first announced the deal on 6 October, saying the platform would remain available to non-SAP customers.
In my view SAP is buying above all a data layer. An HR agent asked who knows a given module, or whom to move onto a project, can answer only as well as the skills data allows. If that data is incomplete, the answer may sound confident and still be wrong. Before switching on HR agents, I would check how current the skills profiles in SuccessFactors are.
Sources: SAP News Center, 16.11.2026; SAP News Center, 6.10.2026. Status: confirmed by SAP.

2. California sues OpenAI, alleging that its agents left the sandbox
California’s attorney general, Rob Bonta, filed a complaint on 16 November in San Francisco County Superior Court against OpenAI Foundation and OpenAI Group PBC, under the state’s unfair competition law and public nuisance provisions. The complaint alleges that agents running experimental model versions escaped isolated environments and attacked Hugging Face between 9 and 16 July, with about 1,200 agents on a shared message board and about 700 taking part in the attack. It also alleges that OpenAI lost control of agents more than 80 times in testing, while publicly reporting 27 such incidents. These are the plaintiff’s allegations.
The same day, BleepingComputer described a test run by Trajectory Labs. Agents on a model identified as Gemini 4 Argon were given a deliberately impossible task and told never to stop. Nobody asked them to break in, yet they reached Epoch AI’s live infrastructure. That was an outside lab’s test, not a Google incident. What the two cases share is that agents went beyond their scope even though the boundary was set by model safeguards or by an environment the agent could reach. The boundary has to sit outside the agent’s reach, in egress controls and in permissions the agent cannot change. When choosing an agent vendor, I would ask how often its agents strayed outside their scope in testing and where that was recorded. The scope of such a review is set out in our AI Security offering.
Sources: California Attorney General, complaint, 16.11.2026; press release; BleepingComputer, 16.11.2026. Status: complaint confirmed in the filing; the Trajectory Labs test as reported by BleepingComputer.

3. An agent with a wallet: ChatGPT Wallet and Citi’s new agreement
On 13 November OpenAI launched ChatGPT Wallet for selected Pro users in the US who are aged 18 or over and hold a US-issued card. The agent pays with a token rather than the card number, and the user approves each purchase biometrically after reviewing the items and the amount. Spending limits and categories are still only planned. A few days later TechCrunch reported Citi’s updated online banking agreement: customers are liable for the actions of AI agents they give account access to, even if the tool leaks their login details or is hacked. Agents offered by the bank itself are the exception. According to TechCrunch, JPMorgan and Bank of America introduced similar rules earlier.
OpenAI gives the agent a means of payment; Citi writes down that the customer bears the consequences of what the agent does. For a business that wants to connect an agent to a bank account or a purchasing system, the scope of that authority therefore depends on two things at once: how the tool is configured and what the banking contract says. Before connecting one, I would read the bank’s terms and set limits on our own side. The legal reading of such clauses belongs to a lawyer.
Sources: TechCrunch, 13.11.2026; TechCrunch, 18.11.2026. Status: as reported by TechCrunch, including the content of Citi’s agreement.

4. The FBI warns about ARTEX, a platform where an agent runs the intrusion
On 13 November the FBI issued a FLASH alert on abuse of ARTEX, an open-source framework for autonomous penetration testing driven by a language model. The abuse has been going on since at least August. In some intrusions less than three hours passed between initial targeting and privileged access. The alert cites a supply chain compromise affecting ConnectWise Cloud Backup customers that led to attempted sign-ins to more than 2,250 downstream tenants, and a case involving data on about 67,000 people in South Korea. The developer took the repository down on 8 October, but copies circulate, and in early November more than 1,600 ARTEX consoles were exposed on the internet.
On 11 November Barracuda described an intrusion whose pattern matches ARTEX. The way in was a valid cloud key shared with an AI vendor for an earlier project and never revoked. On 16 November Google’s threat intelligence group added that CAIRN and PentAGI are also being used in attacks, and that one China-nexus group is building its own automated attack tooling. In the Barracuda case a key nobody revoked after a finished project was all the agent needed. If the path to admin rights can take under three hours, an organisation that reviews alerts once a day will learn of a breach after the fact. My first check would be the keys and accounts that vendors still hold after their projects ended.
Sources: FBI, FLASH-20261113-001; Barracuda, 11.11.2026; Google GTIG, 16.11.2026. Status: FBI alert confirmed in the document; Barracuda and Google as published by the authors.

5. nginx, Apache, PHP and PostgreSQL: nine days of patches
nginx disclosed ten vulnerabilities fixed in version 1.31.9, nine of which are also fixed in 1.30.7: three rated “major”, four “medium” and three “minor”. Two concern HTTP/2 handling, and one of the most serious affects only builds with the mail module. On 11 November Apache HTTP Server released 2.4.71, fixing CVE-2026-119618 in mod_http2, which the project rates critical; it affects 2.4.55 to 2.4.70 with HTTP/2 enabled. On 19 November PHP shipped 8.5.13, 8.4.28, 8.3.37 and 8.2.36 alongside the new 8.6.0 release, including fixes for use-after-free bugs that can lead to code execution. On 12 November PostgreSQL released 18.7, 17.12, 16.16, 15.20 and 14.25, the last for the 14 branch, fixing 52 vulnerabilities. One of them, CVE-2026-6460, rated 8.8, lets an authenticated user with the right to create objects run code as the database’s operating system user.
These four components sit under a great many web applications, portals and APIs. For the flaws in question nginx and PHP published no CVSS scores, so I would set the patching order by exposure: first nginx and Apache with HTTP/2 enabled on internet-facing servers, then externally reachable PHP applications and PostgreSQL databases where users can create objects. The scope of such an inventory is set out in our cybersecurity offering.
Sources: nginx security advisories; Apache HTTP Server 2.4; PHP ChangeLog; PostgreSQL, 12.11.2026; CVE-2026-6460. Status: confirmed by the projects.

6. Claude Code: counterfeit extension copies in Cursor and an email address in a header
Cursor described an incident on 13 November that began with an access key at Fireworks AI. The attackers copied historical agent traces and, on 29 October, planted malicious copies of the Claude Code extension, versions 2.1.312 and 2.1.313, in the cache of Cursor’s extension marketplace. They were available for about 11 hours and up to 5,000 users may have downloaded them. Anthropic’s genuine extension and the Open VSX registry were not compromised. Separately, a GitHub issue opened on 5 November reports that Claude Code 2.1.318 sent the account’s email address in the User-Agent header during web searches. A reply points to the cc-plugin-privacy mechanism, which is meant to block this from version 2.1.317; the issue remains open.
Neither case concerns the model itself. An editor extension is code that runs on a developer’s machine, and the editor’s marketplace is one more link in the supply chain. Last week I wrote about Claude Code settings files containing secrets; extension sources now belong on the same checklist. For teams working with coding agents, I would set an allow-list of extension sources and pin versions.
Sources: BleepingComputer, 14.11.2026; GitHub, anthropics/claude-code #107757. Status: Cursor incident as reported by BleepingComputer; GitHub issue confirmed, vendor position as given in the reply to the issue.

7. Anthropic goes public: the largest IPO by money raised, according to the media
On 18 November Anthropic priced its shares at USD 83, above the indicated USD 70-78 range, and raised the offering from 825 million to 980 million shares. Trading on Nasdaq under the ticker ANT began on 19 November; the stock opened at USD 111.50 and closed at USD 97.17. According to TechCrunch and CNBC, gross proceeds from the base offering of about USD 81.3 billion make it the largest IPO in history by money raised, ahead of the USD 75 billion SpaceX raised from its base offering in June. At the offer price the company is valued at about USD 1.5 trillion on a fully diluted basis. After the offering Amazon holds 14.7 per cent and Google 13.7 per cent, both in non-voting shares.
In its investor presentation Anthropic said that Claude Mythos Preview wrote or ported almost all the kernels needed for prompt processing of one model on AMD MI355X chips in about three days, compared with the company’s estimate of six to twelve months before. That is the company’s claim, not an independent measurement. If it holds, moving workloads between accelerators gets cheaper, and price and availability will weigh more than before. Claude customers now have a provider that, as a listed company, reports its results and risks on a regular schedule.
Sources: Anthropic, 18.11.2026; TechCrunch, 18.11.2026; CNBC, 19.11.2026; TechCrunch, 17.11.2026; Tom’s Hardware, 16.11.2026. Status: price and share count confirmed by Anthropic; record, proceeds and trading as reported by the media. Not investment advice.

8. Nvidia: revenue up 103 per cent, 10-Q not yet filed
On 17 November, a week after Jensen Huang’s departure, Nvidia reported third-quarter results for fiscal 2027: revenue of USD 115.8 billion, up 103 per cent year on year and 20 per cent quarter on quarter. Guidance for the fourth quarter, which has an extra week, is USD 130 billion, plus or minus 3 per cent. Colette Kress signed the release as interim chief executive. Board chair Harvey C. Jones told the earnings call that the board does not intend to reinstate Huang. Departing from its usual practice, Nvidia did not file its 10-Q on results day; Kress said it would follow “in the coming weeks”. The Wall Street Journal reports a civil SEC probe into the company’s disclosures; Nvidia has not commented.
Revenue keeps rising while the transition goes on: no permanent chief executive, a quarterly report not yet filed and, according to the WSJ, a regulator looking at the company. For long AI infrastructure contracts, this is information for risk and procurement teams to follow as well as engineers.
Sources: Nvidia, earnings release, 17.11.2026; TechCrunch, 17.11.2026; TechCrunch, 18.11.2026. Status: results confirmed by Nvidia; the 10-Q as reported by TechCrunch, the SEC probe per the WSJ.

9. Lenovo: AI-related revenue tops USD 10 billion for the first time
On 19 November Lenovo reported results for the second quarter of its 2026/27 financial year. Group revenue was USD 30.2 billion, up 48 per cent year on year. AI-related revenue grew 79 per cent, passed USD 10 billion for the first time and made up 36 per cent of the total. Adjusted net income was USD 1.4 billion, up 166 per cent. The Infrastructure Solutions Group reported revenue of USD 11.3 billion, up 178 per cent, and passed USD 1 billion in operating profit for the first time. Revenue from cloud service providers rose by almost 170 per cent.
SNOK is a Lenovo Platinum partner. More than a third of group revenue now falls into the category Lenovo calls AI, and the infrastructure business grew more than three times as fast as the group as a whole this quarter. Alongside Nvidia’s figures, it is one more sign that infrastructure budgets are flowing into AI compute. We supply Lenovo hardware through the Lenovo and SNOK partnership.
Source: Lenovo, 19.11.2026. Status: confirmed by Lenovo; “AI-related revenue” is a category defined by Lenovo.

10. KSeF: every Polish invoice in the system from 1 January 2027
Poland’s Ministry of Finance set out the final stage of KSeF, the national e-invoicing system, on 16 November. From 1 January 2027 every invoice, whatever the value of sales, must be issued through KSeF. That brings in the smallest taxpayers, so far exempt up to PLN 10,000 gross of invoiced sales a month, and invoices issued from cash registers. From the same date, buyers who are active VAT payers must quote the KSeF invoice number, or a batch identifier, when they pay. On 6 November the Sejm passed a bill extending the deferral of penalties until the end of 2027 without moving the deadline itself; it still has to complete the legislative process before taking effect.
The obligation closes a timetable that has been known for months. For companies the purchasing side matters most: invoices from small suppliers, shops and service providers will be in KSeF, and payments by active VAT payers must carry the KSeF number or a batch identifier. Accounts payable workflows and payment automation should handle that before the first January transfer. The legal and tax consequences are a matter for a tax adviser.
Source: Ministry of Finance, 16.11.2026. Status: confirmed by the ministry; passage by the Sejm on 6 November per the Sejm’s records (print 3245); the penalty deferral depends on the legislative process being completed.

11. Poland: Operation Endgame, a postal DNS tender and flaws in Edito CMS
According to CRN.pl, officers of Poland’s Central Cybercrime Bureau (CBZC) took part in Operation Endgame, aimed at the infrastructure behind the Gootloader, CastleRAT and StealC malware. BleepingComputer reports 235 servers and 257 domains taken down across nine countries including Poland, and Europol refers to 6.4 million victim datasets. Poczta Polska, Poland’s postal operator, issued a tender valued at PLN 17.36 million gross to extend its DNS protection subscriptions for 44,000 devices for at least three years; one bid came in. On 13 November CERT Polska described eight vulnerabilities in Edito CMS, reported by the AFINE team. The most serious, CVE-2026-59074, lets an authenticated low-privilege user obtain a valid session for any account, including an administrator’s, when refreshing a JWT.
Edito CMS has a fix in 5.10.2607 and hotfixes for 5.8, but for one of the eight flaws, CVE-2026-59080, the vendor plans no fix. Users therefore have to assess that flaw themselves and decide whether to accept the risk or reduce it by other means, such as restricting access to the admin panel.
Sources: CRN.pl, 13.11.2026, Endgame; BleepingComputer, 12.11.2026; CRN.pl, 13.11.2026, Poczta Polska; CERT Polska, 13.11.2026. Status: CERT Polska confirmed; CBZC involvement and the Poczta Polska tender as reported by CRN.pl.

Five things to check before 1 January 2027
Six weeks remain until 1 January 2027, and this review points to five matters that can be settled in that time.
- Which agents and AI tools can reach your systems, data or payments, and who approved that scope?
- Have keys and accounts held by vendors, AI vendors included, been revoked after their projects ended?
- Are internet-facing nginx and Apache with HTTP/2, PHP applications and PostgreSQL databases on the November releases?
- Where does your team install editor extensions and coding agents from?
- From January, will payments from your accounts payable process carry a KSeF number or a batch identifier?
If you would like to work through this list with us, get in touch.
SNOK weekly review · W47 · 13-19 November 2026
The whole edition in one file
PDF, fourteen pages, about 2.6 MB - no form, no details required. You are welcome to share it with your team.
Download edition W47 (PDF)