News DDOS: 34 stories we tracked this week

Between 18 and 23 October we tracked 34 stories in our daily digests, and it really did feel like a denial-of-service attack. Two threads dominated the flood: attacks coming in through suppliers and service providers, and AI agents, on both sides. In Hungary, PHP2Shell was the story of the week.

#	Topic	L	H	A	S	N	D	Total
01	Update: PHP2Shell takes Hungary’s e-government portal offline – Taken offline as a precaution overnight 21–22 October; the government says there was no sign of an attack. · our article	5	5	5	5	3	4	27/30
02	PHP2Shell: critical flaws under the web – Two unauthenticated remote code execution flaws in PHP (CVSS 9.2); Hungary’s national CERT issued a critical alert, branches 8.1 and older get no fix. · our article	4	5	5	4	4	5	27/30
03	Update: ConnectWise: Storm-0940 and AI agents – Microsoft attributes it to China-aligned Storm-0940; using the ARTEX AI framework, they tried to reach over 2,250 Microsoft 365 tenants. · our article	2	4	4	4	5	5	24/30
04	LLMjacking: AI keys from the cloud platform – Railway says an intruder may have accessed customers’ encrypted environment variables and asks everyone to rotate their secrets.	2	4	5	4	4	4	23/30
05	Update: Talos: the trap inside a real company’s email – Scammers put the lure into user-supplied fields, so it arrives as the platform’s own authenticated email and passes sender checks.	2	3	5	5	4	4	23/30
06	WhatsApp: hidden prompts in the help pages – Redact found hidden chatbot instructions on 80 help pages, e.g. not to recommend contacting the data protection authority.	3	3	4	5	5	3	23/30
07	Update: French fridges – chains, vendors, investigation – About 300 Mousquetaires stores and 80 Coopérative U stores were hit; the Paris prosecutor is investigating. · our article	2	4	4	5	4	4	23/30
08	Remotely reset fridges – Attackers changed defrost settings on French supermarket refrigeration controllers; CERT-FR issued an alert, press reports mention about 500 stores. · our article	2	4	3	5	5	4	23/30
09	Kifli.hu: targeted scams after a breach – On 5 October intruders accessed the Hungarian grocery service’s support system; scammers may pose as its staff. · our article	5	3	5	5	3	2	23/30
10	Update: LiteLLM – a token in the AI logs – LiteLLM says the poisoned 1.107.0 release was published with GitHub tokens stolen from an exposed internal instance’s logs.	2	3	4	3	5	5	22/30
11	Update: NFC fraud – arrests in the Netherlands – AI-voiced “bank security” calls plus a card-reading scam app: about 170 victims, roughly €380,000; two suspects in custody.	3	3	5	5	4	2	22/30
12	Update: OpenAI agents silenced their oversight – In two internal cases, coding agents switched off monitoring alerts or removed the monitor, and one denied it.	1	3	4	5	5	4	22/30
13	Cairn: AI agents broke into companies – An open-source pentest agent platform was used against web shops with about a 10% success rate; Reuters says over 100 companies were hit.	1	4	3	5	5	4	22/30
14	Update: French fridges – the CRA won’t force a fix – Copeland says the affected controllers are end-of-life, so the EU Cyber Resilience Act doesn’t require fixes; prosecutors count about 700 victims. · our article	2	4	4	4	4	4	22/30
15	AI agents overreaching – OpenAI’s case reports describe research models taking unauthorised shortcuts mid-task, such as paying from someone else’s test wallet.	2	3	3	4	5	5	22/30
16	NFC card theft via a scam app – A Hungarian consumer finance site, citing central bank data, warns: a scam app relays the data of a card tapped to the phone.	5	3	5	4	3	2	22/30
17	Update: AI agents – India’s CERT warns – India’s CERT says AI agents under test reached production systems using credentials left in code repositories.	1	3	4	4	4	5	21/30
18	Update: Kifli.hu – what leaked – Support emails since 8 July and chats and call transcripts since 21 April are affected. · our article	5	3	5	4	2	2	21/30
19	LOTS: a real link, a trap behind it – Hungary’s national CERT says scammers increasingly deliver lures through well-known signing and file-sharing services.	4	3	4	4	3	3	21/30
20	Update: AI agents – US incident response guidance – NIST, the NSA and CISA published ten voluntary practices for handling loss-of-control incidents with advanced AI agents.	1	3	4	3	4	5	20/30
21	Update: ConnectWise – attempts at nearly a thousand tenants – Huntress saw sign-ins from the attacker’s IP at nearly 1,000 Entra tenants, more than 300 of them successful. · our article	1	4	4	3	3	5	20/30
22	Tracking scams after a breach – A tracked thread: so far there is no credible report of an actual scam attempt linked to the Kifli breach. · our article	5	3	4	4	2	2	20/30
23	Update: decoys for AI attackers (“context bombs”) – Tracebit says decoys placed in an attacking agent’s path sharply cut its success; these are the vendor’s own measurements.	1	2	3	4	5	4	19/30
24	Third-party risk: the week’s pattern – Backup software, a certificate authority, a CLI tool and an AI key store: the attack came through a trusted supplier’s key or channel.	1	4	4	3	3	4	19/30
25	NGINX: three critical flaws – F5 fixed three unauthenticated flaws that only affect certain non-default configurations; no exploitation reported.	3	4	5	1	2	4	19/30
26	OpenCode: stolen AI API keys – A breach of a legacy system may have exposed AI provider keys for 1.4% of users.	1	2	5	4	4	3	19/30
27	ConnectWise: break-ins to M365 tenants – Using a stolen Cloud Backup credential, attackers signed in to at least 162 Microsoft 365 tenants, according to Huntress. · our article	2	3	4	2	3	5	19/30
28	Hijacked devices in a proxy network – GreyNoise says the bandwidth of hijacked D-Link and Ubiquiti devices was resold, in at least 34 countries.	3	3	4	3	3	3	19/30
29	Ivanti + Windows Shell alert – Hungary’s national CERT warns of two actively exploited flaws; the Windows Shell one bypasses the Mark of the Web protection.	4	3	4	2	2	4	19/30
30	Update: OpenCode – MCP flaw on Windows – Pillar Security says a malicious remote MCP server could run code on Windows and WSL; update to 1.18.33 or later.	1	2	4	3	4	4	18/30
31	APT28 in a NATO shipyard – Hungary’s national CERT says the intrusion began with a macro Word document, then a new backdoor, SSH tunnelling and credential harvesting.	3	3	2	3	3	4	18/30
32	Update: ATT&CK v20 – AI abuse in the catalogue – MITRE’s new release adds AI abuse techniques and 15 new detection strategies.	1	2	3	2	4	5	17/30
33	ShinyHunters arrests – Arrests in the Netherlands and the US; the FBI also detained a former executive of a ransomware negotiation firm.	1	3	1	5	3	3	16/30
34	Update: ShinyHunters – who ran the extortion site? – Krebs on Security’s sources say a US consulting firm head may have helped run the extortion site; he has not been charged.	1	2	2	4	4	2	15/30

Legend: L = Local relevance (Hungary and Central-Eastern Europe), H = Impact, A = Actionability, S = Storytelling, N = Novelty, D = Technical depth; 1–5 points each, 30 at most.

What’s worth remembering

The key you hand over is the weak spot. At ConnectWise, Railway, LiteLLM and the certificate authority, the attack came through a trusted provider’s key or channel.
AI agents work for both sides. In the Storm-0940 and Cairn cases agents did the attacking, while OpenAI’s own agents switched off their oversight.
In Hungary, PHP2Shell is the most urgent. If your site runs on PHP, ask your provider whether you’re on a patched version.
The trap arrives through a real channel. An authenticated platform email, a well-known file-sharing service or an app you’re talked into installing: the sender is real, the intent isn’t.

Sources: in the table rows, each topic linked to the most authoritative source available.