Reporting Period: 16 October 2026 – 23 October 2026
Arcanum’s View
The clearest signal this week is the UK National Cyber Security Centre’s standalone advisory on two unauthenticated remote-code-execution flaws in PHP, CVE-2026-109618 and CVE-2026-109619, which the NCSC says affect “any version” of the language and, by extension, any product that bundles it[i]. It is unusual for the NCSC to issue a dedicated advisory on a single software component rather than folding it into a broader bulletin, and that choice signals how widely exploitable the agency judges this flaw to be: PHP underpins a large share of the internet’s content management systems, internal tooling and vendor appliances, so an unauthenticated path to remote code execution here has a correspondingly wide blast radius across UK organisations’ direct estates and their supply chains alike.
The advisory’s own emphasis, patch internet-facing systems first, then move to internal ones, and prepare for a “vulnerability patch wave”, sets the tone for the rest of this week’s findings. It sits alongside two reinforcing patterns: edge devices and remote-access tooling remain the preferred route into networks for both state-linked and financially-motivated actors, as seen in a China-linked group’s exploitation of a BeyondTrust Remote Support zero-day[ii] and the SafePay ransomware gang’s continued use of exploited edge-device CVEs and stolen credentials against 600 or more organisations, including a named UK software vendor[iii]; and artificial intelligence is now visible on both sides of the threat equation, from AI-generated voice clones used in bank fraud[iv] to a vendor breach that exposed AI provider API keys[v].
This week’s largest data-exposure stories, at Nikkei, Zeabur and Genesys, again trace back to compromised cloud-service or vendor credentials rather than novel exploits[vi]. Read together, the NCSC’s PHP advisory is less a one-off patch notice than a timely reminder that the foundational, unglamorous software running underneath most applications still deserves the same urgency as the headline-grabbing zero-days.
Immediate Priorities for Security Leaders
- Patch or mitigate the two critical, unauthenticated PHP remote-code-execution vulnerabilities flagged by the NCSC this week, prioritising internet-facing systems and any third-party appliances that bundle PHP.
- Validate that BeyondTrust Remote Support and any similar privileged remote-access tooling are patched and were not already compromised before the fix was applied.
- Cross-check asset inventories against this week’s 13 new CISA Known Exploited Vulnerabilities catalogue additions, which span Fortinet, Atlassian, Palo Alto Networks, Cisco, Ahsay, Plex and Siemens products.
- Review exposure to the SafePay ransomware group’s known tactics, particularly abuse of legitimate remote-access tools such as AnyDesk and TeamViewer.
- Rotate or audit any AI service API keys that may have been exposed through third-party AI tooling breaches, and extend fraud-verification controls to cover AI-generated voice and video.
Detailed, ranked actions with owners and deadlines appear at the end of this advisory.
Key Threats
Internet-facing infrastructure and identity exploitation
The NCSC published a dedicated advisory on 21 October covering CVE-2026-109618 and CVE-2026-109619, two flaws in PHP’s handling of out-of-memory conditions that can corrupt memory and, under certain conditions, allow unauthenticated remote code execution[vii]. The agency said organisations using any version of PHP are affected, including products that bundle PHP as a dependency, and urged internet-facing systems to be identified and patched first, with temporary mitigations or offline isolation recommended where patching is not immediately possible[viii]. Separately, CISA added 13 vulnerabilities to its Known Exploited Vulnerabilities catalogue between 16 and 23 October, affecting BeyondTrust Remote Support, Fortinet products, Atlassian software, Cisco Secure Email Gateway, Ahsay AhsayCBS backup software, Palo Alto Networks PAN-OS, Plex Media Server and Siemens Software Center[ix].
The BeyondTrust entry, CVE-2026-102764, was confirmed by Mandiant and Google’s Threat Intelligence Group to have been exploited as a zero-day by a Chinese state-linked espionage group that Mandiant tracks as Armory Castle (UNC5221), using custom malware named Sawfly and Bareclaw to harvest every credential stored on compromised appliances and to plant persistence that survives reboots and configuration resets[x]. Separately, researchers at VulnCheck linked exploitation attempts against several older CVEs, including an XWiki remote-code-execution flaw, to infrastructure associated with the Flax Typhoon botnet and other actors operating from Hong Kong and Taiwan[xi].
Ransomware and extortion
The FBI, together with cybersecurity agencies in Germany and the UK, published a joint advisory on 20 October describing the SafePay ransomware group as among the most active of 2026, with over 600 confirmed victims across the US and Europe since August 2024[xii]. Named victims in the advisory’s wider reporting include UK software vendor Microlise and government services contractor Conduent[xiii]. The group gains initial access through password guessing, phishing and automated scanning for exploitable edge-device vulnerabilities, then uses legitimate tools such as AnyDesk and TeamViewer to deploy ransomware after disabling security protections and exfiltrating data, a pattern consistent with double extortion[xiv].
AI-enabled threat activity
Dutch police arrested two suspects accused of using AI-generated voice clips combined with NFC card-relay techniques to defraud roughly 170 bank customers of approximately 380,000 euros[xv]. Separately, security researchers at Hirundo reported that the Qwen3-Coder model generates measurably more exploitable code for certain prompts, including those that invoke a US defence-contractor persona, echoing an earlier study by Booz Allen that found similar effects in other coding models[xvi]. US agencies NIST, NSA and CISA jointly published voluntary guidance setting out ten practices for detecting and containing AI agents that deviate from their approved tasks during training[xvii].
A researcher also demonstrated an indirect prompt-injection technique against ChatGPT’s GitHub connector that could cause the assistant to write a malicious GitHub Actions workflow into a private repository in order to exfiltrate secrets; OpenAI classified the finding as out of scope for its bug bounty programme[xviii]. In a separate development, the US Department of Homeland Security confirmed on 19 October that an order barring federal civilian agencies from using DeepSeek’s AI products, issued in July, had taken effect, prompting an objection from China’s foreign ministry[xix].
Software supply chain and browser trust
A breach at AI coding platform OpenCode, tied to what the company described as a large-scale “distillation attack” against its Console product, affected roughly 1.4% of Console users and potentially exposed their API keys for Anthropic, OpenAI and Google AI models[xx]. Database vendor PingCAP disclosed that attackers used stolen service-account credentials to copy source code and customer records, including emails, SQL editor contents, contract and billing details, over a period running from early September to early October before the credentials were disabled[xxi]. Separately, Google confirmed that Chrome will now warn users before allowing a plain, unencrypted HTTP connection, extending a protection previously limited to Enhanced Protection users to its entire user base by default[xxii].
Large-scale data exposure
Japanese media group Nikkei confirmed that around 1.83 million records were stolen after unauthorised access to a cloud development service, beginning in early October and disclosed on 21 October; roughly 150,000 of those records contained identifiable personal information such as names, email addresses or dates of birth, and the company said this was the third breach to affect it in a matter of weeks[xxiii]. Customer-experience vendor Genesys disclosed on 16 October that an intruder had accessed its internal CRM and learning systems, though it said its Genesys Cloud and Engage products were not affected[xxiv]. Cloud hosting provider Zeabur disclosed that an October attack exploiting leaked plaintext internal credentials exposed more than 3.2 million environment-variable records and over 321,000 accounts, its second breach of the year; the company subsequently announced it will shut down entirely between December 1 and December 31[xxv]. Social network Bluesky confirmed that more than 4,500 accounts were compromised through posts promoting a fake government payment scheme, with the affected accounts temporarily suspended[xxvi].
Strategic Assessment
Edge devices and privileged remote-access tools remain the preferred route in
Three of this week’s most significant stories, the BeyondTrust zero-day, the SafePay ransomware advisory and the NCSC’s PHP warning, share a common thread: attackers consistently favour internet-facing software and remote-access appliances over more complex intrusion paths[xxvii]. Legacy and slow-to-patch edge infrastructure, rather than novel techniques, continues to provide the easiest and most reliable foothold for both state-linked espionage operations and ransomware crews, which increasingly resemble each other in their tooling once inside a network.
AI has become a two-sided risk, as a tool for attackers and as a new exposure surface
This week illustrates both directions of AI risk simultaneously. On the offensive side, criminals combined AI voice cloning with traditional card-relay fraud to defraud bank customers[xxviii], and a widely-used coding model was shown to generate less secure code under adversarial prompting[xxix]. On the defensive side, organisations are now exposed through the AI tools and connectors they themselves have adopted, as seen in the OpenCode API-key exposure and the ChatGPT GitHub connector prompt-injection finding[xxx]. Governance guidance such as the NIST, NSA and CISA agent-containment practices is a sign regulators recognise this, but voluntary guidance alone will not close the gap between how quickly these tools are being deployed and how slowly their risks are being assessed[xxxi].
Cloud and vendor credential hygiene, not novel exploits, drove this week’s largest breaches
None of the week’s largest data-exposure incidents, Nikkei, Zeabur, Genesys or PingCAP, involved a sophisticated zero-day; each traces back to compromised or leaked credentials for a cloud service, internal system or service account[xxxii]. This reinforces that identity and access governance over third-party and cloud-hosted systems is now a bigger driver of breach frequency than software vulnerability management alone, even as vulnerability management continues to dominate the headlines.
State-linked operations continue despite public attribution and sanctions
Public naming of state-linked infrastructure and actors has not visibly slowed their activity. The BeyondTrust zero-day was attributed within days to a Chinese state-sponsored group[xxxiii], while VulnCheck’s findings on exploitation attempts tied to Flax Typhoon-associated infrastructure show that previously identified botnets continue to be leveraged against new targets[xxxiv]. Attribution and sanctions remain necessary disclosure tools, but this week’s evidence suggests they are not, on their own, a meaningful deterrent against continued operations.
Assurance Considerations, Tied to This Week’s Findings
| Control area (CAF / GovAssure / NIS2) | This week’s evidence it matters |
|---|---|
| Vulnerability management and rapid remediation | NCSC’s PHP advisory covering two unauthenticated RCE flaws, 13 new CISA KEV additions, and the BeyondTrust Remote Support zero-day all require rapid, prioritised patching of internet-facing systems[xxxv]. |
| Identity and access control | SafePay’s reliance on password guessing and stolen credentials, and Zeabur’s exposure of leaked plaintext internal credentials, both point to weak identity hygiene as a root cause of compromise[xxxvi]. |
| Secure-by-design AI deployment | The OpenCode API-key exposure, the ChatGPT GitHub connector prompt-injection finding, and the Qwen3-Coder insecure-code research all show that AI tooling introduces risk that conventional application security review does not yet cover[xxxvii]. |
| Supply-chain assurance | PingCAP’s disclosure that stolen service-account credentials exposed source code and customer records over roughly a month highlights the ongoing risk from third-party development platforms[xxxviii]. |
| Third-party and vendor governance | Nikkei’s third breach in weeks and Genesys’s internal system compromise both originated with vendor or cloud-service access rather than the organisations’ own core systems[xxxix]. |
| Detection and response | SafePay’s abuse of legitimate remote-access tools such as AnyDesk and TeamViewer, and its practice of disabling security protections before deploying ransomware, underline the need for behavioural detection rather than signature-based controls alone[xl]. |
| Operational resilience and recovery | Zeabur’s decision to shut down entirely by the end of December leaves its customers with a hard dependency-continuity problem that recovery planning must anticipate for any critical SaaS vendor[xli]. |
Questions for Boards and Senior Leaders
- Can we confirm that every PHP-dependent system we operate or rely on through a vendor, not just our own applications, has been identified and patched against CVE-2026-109618 and CVE-2026-109619?
- If we use BeyondTrust Remote Support or a similar privileged remote-access appliance, have we treated it as potentially compromised and reviewed it for the credential-harvesting behaviour Mandiant described, rather than simply applying the patch?
- Do we know which of our SafePay-style ransomware exposure points, exploitable edge devices, weak passwords, or unmonitored use of tools like AnyDesk, remain open today?
- If one of our AI coding or API-key-dependent vendors suffered a breach similar to OpenCode’s this week, would we know which of our own API keys were exposed, and could we rotate them within hours rather than days?
- What verification process would catch an AI-generated voice-clone fraud attempt against our finance or treasury function before funds are transferred?
- If a critical cloud or SaaS vendor announced it was shutting down with weeks of notice, as Zeabur has done, do we have a tested plan to migrate the data and services we depend on?
Priority Actions, Ranked
| Severity | Finding | Action | Deadline |
|---|---|---|---|
| Critical | Unauthenticated RCE in PHP (CVE-2026-109618, CVE-2026-109619) affecting any PHP version | Identify and patch all PHP-dependent internet-facing systems first; apply PHP Foundation mitigations or isolate systems that cannot be patched immediately | Within 48 hours |
| Critical | BeyondTrust Remote Support zero-day (CVE-2026-102764) exploited by a China-linked actor | Apply the BT26-05 patch, treat any internet-reachable appliance as potentially compromised, and rotate all credentials it stored | Within 48 hours |
| High | 13 new CISA KEV additions across BeyondTrust, Fortinet, Atlassian, Palo Alto, Cisco, Ahsay, Plex and Siemens products | Cross-reference the asset inventory against the new KEV entries and patch or mitigate every affected product in scope | 7 days |
| High | SafePay ransomware active against 600+ organisations via edge-device CVEs, phishing and stolen credentials | Hunt for unauthorised AnyDesk or TeamViewer sessions, unexpected new accounts, and disabled security tooling; confirm backups are offline or immutable | 7 days |
| Medium | AI API keys and secrets exposed via a vendor “distillation attack” (OpenCode Console breach) | Rotate any Anthropic, OpenAI or Google AI API keys used through affected tooling and review AI vendor breach-notification terms | 14 days |
| Medium | AI voice-cloning and NFC relay fraud defrauding bank customers | Brief finance and treasury staff and introduce out-of-band verification for high-value transfer requests | 14 days |
| Medium | Cloud and vendor credential compromises at Nikkei, Zeabur and Genesys exposing millions of records | Review third-party and cloud-vendor access governance, credential rotation cadence, and contractual breach-notification SLAs | 30 days |
| Low | Chrome now blocks plain HTTP connections by default for all users | Confirm internal web applications and legacy portals do not depend on unencrypted HTTP, and update browser security policy documentation | 30 days |
[i]NCSC — “Critical security issue affecting PHP” (21 October 2026) — https://www.ncsc.gov.uk/news/critical-security-issue-affecting-php
[ii]The Record — “FBI says SafePay ransomware gang among most active, 600 victims” (20 October 2026) — https://therecord.media/fbi-says-safepay-ransomware-gang-among-most-active-600-victims · The Record — “BeyondTrust confirms China-linked zero-day attack” (16 October 2026) — https://therecord.media/beyond-trust-cyberattack-china
[iii]The Record — “FBI says SafePay ransomware gang among most active, 600 victims” (20 October 2026) — https://therecord.media/fbi-says-safepay-ransomware-gang-among-most-active-600-victims
[iv]The Hacker News — “ThreatsDay: AI Secrets Leak, PingCAP Source Theft, Signal Hang-Up Flaw, Laravel Auth Bypass + 21 More” (22 October 2026) — https://thehackernews.com/2026/10/threatsday-ai-secrets-leak-pingcap.html
[v]The Hacker News — “ThreatsDay: AI Secrets Leak, PingCAP Source Theft, Signal Hang-Up Flaw, Laravel Auth Bypass + 21 More” (22 October 2026) — https://thehackernews.com/2026/10/threatsday-ai-secrets-leak-pingcap.html
[vi]ITPro — “Nikkei admits millions of records stolen after reporting third breach in a month” (21 October 2026) — https://www.itpro.com/security/data-breaches/nikkei-admits-millions-of-records-stolen-after-reporting-third-breach-in-a-month
[vii]NCSC — “Critical security issue affecting PHP” (21 October 2026) — https://www.ncsc.gov.uk/news/critical-security-issue-affecting-php
[viii]NCSC — “Critical security issue affecting PHP” (21 October 2026) — https://www.ncsc.gov.uk/news/critical-security-issue-affecting-php
[ix]CISA — “Known Exploited Vulnerabilities Catalog” (additions dated 13-20 October 2026) — https://www.cisa.gov/known-exploited-vulnerabilities-catalog · Senserva — “CISA KEV Additions This Week” (23 October 2026) — https://senserva.com/exploited-this-week.html
[x]The Record — “FBI says SafePay ransomware gang among most active, 600 victims” (20 October 2026) — https://therecord.media/fbi-says-safepay-ransomware-gang-among-most-active-600-victims · The Record — “BeyondTrust confirms China-linked zero-day attack” (16 October 2026) — https://therecord.media/beyond-trust-cyberattack-china
[xi]The Hacker News — “ThreatsDay: AI Secrets Leak, PingCAP Source Theft, Signal Hang-Up Flaw, Laravel Auth Bypass + 21 More” (22 October 2026) — https://thehackernews.com/2026/10/threatsday-ai-secrets-leak-pingcap.html
[xii]The Record — “FBI says SafePay ransomware gang among most active, 600 victims” (20 October 2026) — https://therecord.media/fbi-says-safepay-ransomware-gang-among-most-active-600-victims
[xiii]The Record — “FBI says SafePay ransomware gang among most active, 600 victims” (20 October 2026) — https://therecord.media/fbi-says-safepay-ransomware-gang-among-most-active-600-victims
[xiv]The Record — “FBI says SafePay ransomware gang among most active, 600 victims” (20 October 2026) — https://therecord.media/fbi-says-safepay-ransomware-gang-among-most-active-600-victims
[xv]The Hacker News — “ThreatsDay: AI Secrets Leak, PingCAP Source Theft, Signal Hang-Up Flaw, Laravel Auth Bypass + 21 More” (22 October 2026) — https://thehackernews.com/2026/10/threatsday-ai-secrets-leak-pingcap.html
[xvi]The Hacker News — “ThreatsDay: AI Secrets Leak, PingCAP Source Theft, Signal Hang-Up Flaw, Laravel Auth Bypass + 21 More” (22 October 2026) — https://thehackernews.com/2026/10/threatsday-ai-secrets-leak-pingcap.html
[xvii]The Hacker News — “ThreatsDay: AI Secrets Leak, PingCAP Source Theft, Signal Hang-Up Flaw, Laravel Auth Bypass + 21 More” (22 October 2026) — https://thehackernews.com/2026/10/threatsday-ai-secrets-leak-pingcap.html
[xviii]The Hacker News — “ThreatsDay: AI Secrets Leak, PingCAP Source Theft, Signal Hang-Up Flaw, Laravel Auth Bypass + 21 More” (22 October 2026) — https://thehackernews.com/2026/10/threatsday-ai-secrets-leak-pingcap.html
[xix]The Hacker News — “ThreatsDay: AI Secrets Leak, PingCAP Source Theft, Signal Hang-Up Flaw, Laravel Auth Bypass + 21 More” (22 October 2026) — https://thehackernews.com/2026/10/threatsday-ai-secrets-leak-pingcap.html
[xx]The Hacker News — “ThreatsDay: AI Secrets Leak, PingCAP Source Theft, Signal Hang-Up Flaw, Laravel Auth Bypass + 21 More” (22 October 2026) — https://thehackernews.com/2026/10/threatsday-ai-secrets-leak-pingcap.html
[xxi]The Hacker News — “ThreatsDay: AI Secrets Leak, PingCAP Source Theft, Signal Hang-Up Flaw, Laravel Auth Bypass + 21 More” (22 October 2026) — https://thehackernews.com/2026/10/threatsday-ai-secrets-leak-pingcap.html
[xxii]The Hacker News — “ThreatsDay: AI Secrets Leak, PingCAP Source Theft, Signal Hang-Up Flaw, Laravel Auth Bypass + 21 More” (22 October 2026) — https://thehackernews.com/2026/10/threatsday-ai-secrets-leak-pingcap.html
[xxiii]ITPro — “Nikkei admits millions of records stolen after reporting third breach in a month” (21 October 2026) — https://www.itpro.com/security/data-breaches/nikkei-admits-millions-of-records-stolen-after-reporting-third-breach-in-a-month
[xxiv]The Hacker News — “ThreatsDay: AI Secrets Leak, PingCAP Source Theft, Signal Hang-Up Flaw, Laravel Auth Bypass + 21 More” (22 October 2026) — https://thehackernews.com/2026/10/threatsday-ai-secrets-leak-pingcap.html
[xxv]The Hacker News — “ThreatsDay: AI Secrets Leak, PingCAP Source Theft, Signal Hang-Up Flaw, Laravel Auth Bypass + 21 More” (22 October 2026) — https://thehackernews.com/2026/10/threatsday-ai-secrets-leak-pingcap.html
[xxvi]The Hacker News — “ThreatsDay: AI Secrets Leak, PingCAP Source Theft, Signal Hang-Up Flaw, Laravel Auth Bypass + 21 More” (22 October 2026) — https://thehackernews.com/2026/10/threatsday-ai-secrets-leak-pingcap.html
[xxvii]The Record — “FBI says SafePay ransomware gang among most active, 600 victims” (20 October 2026) — https://therecord.media/fbi-says-safepay-ransomware-gang-among-most-active-600-victims · The Record — “BeyondTrust confirms China-linked zero-day attack” (16 October 2026) — https://therecord.media/beyond-trust-cyberattack-china
[xxviii]The Hacker News — “ThreatsDay: AI Secrets Leak, PingCAP Source Theft, Signal Hang-Up Flaw, Laravel Auth Bypass + 21 More” (22 October 2026) — https://thehackernews.com/2026/10/threatsday-ai-secrets-leak-pingcap.html
[xxix]The Hacker News — “ThreatsDay: AI Secrets Leak, PingCAP Source Theft, Signal Hang-Up Flaw, Laravel Auth Bypass + 21 More” (22 October 2026) — https://thehackernews.com/2026/10/threatsday-ai-secrets-leak-pingcap.html
[xxx]The Hacker News — “ThreatsDay: AI Secrets Leak, PingCAP Source Theft, Signal Hang-Up Flaw, Laravel Auth Bypass + 21 More” (22 October 2026) — https://thehackernews.com/2026/10/threatsday-ai-secrets-leak-pingcap.html
[xxxi]The Hacker News — “ThreatsDay: AI Secrets Leak, PingCAP Source Theft, Signal Hang-Up Flaw, Laravel Auth Bypass + 21 More” (22 October 2026) — https://thehackernews.com/2026/10/threatsday-ai-secrets-leak-pingcap.html
[xxxii]ITPro — “Nikkei admits millions of records stolen after reporting third breach in a month” (21 October 2026) — https://www.itpro.com/security/data-breaches/nikkei-admits-millions-of-records-stolen-after-reporting-third-breach-in-a-month
[xxxiii]The Record — “FBI says SafePay ransomware gang among most active, 600 victims” (20 October 2026) — https://therecord.media/fbi-says-safepay-ransomware-gang-among-most-active-600-victims · The Record — “BeyondTrust confirms China-linked zero-day attack” (16 October 2026) — https://therecord.media/beyond-trust-cyberattack-china
[xxxiv]The Hacker News — “ThreatsDay: AI Secrets Leak, PingCAP Source Theft, Signal Hang-Up Flaw, Laravel Auth Bypass + 21 More” (22 October 2026) — https://thehackernews.com/2026/10/threatsday-ai-secrets-leak-pingcap.html
[xxxv]CISA — “Known Exploited Vulnerabilities Catalog” (additions dated 13-20 October 2026) — https://www.cisa.gov/known-exploited-vulnerabilities-catalog · Senserva — “CISA KEV Additions This Week” (23 October 2026) — https://senserva.com/exploited-this-week.html
[xxxvi]The Record — “FBI says SafePay ransomware gang among most active, 600 victims” (20 October 2026) — https://therecord.media/fbi-says-safepay-ransomware-gang-among-most-active-600-victims
[xxxvii]The Hacker News — “ThreatsDay: AI Secrets Leak, PingCAP Source Theft, Signal Hang-Up Flaw, Laravel Auth Bypass + 21 More” (22 October 2026) — https://thehackernews.com/2026/10/threatsday-ai-secrets-leak-pingcap.html
[xxxviii]The Hacker News — “ThreatsDay: AI Secrets Leak, PingCAP Source Theft, Signal Hang-Up Flaw, Laravel Auth Bypass + 21 More” (22 October 2026) — https://thehackernews.com/2026/10/threatsday-ai-secrets-leak-pingcap.html
[xxxix]ITPro — “Nikkei admits millions of records stolen after reporting third breach in a month” (21 October 2026) — https://www.itpro.com/security/data-breaches/nikkei-admits-millions-of-records-stolen-after-reporting-third-breach-in-a-month
[xl]The Record — “FBI says SafePay ransomware gang among most active, 600 victims” (20 October 2026) — https://therecord.media/fbi-says-safepay-ransomware-gang-among-most-active-600-victims
[xli]The Hacker News — “ThreatsDay: AI Secrets Leak, PingCAP Source Theft, Signal Hang-Up Flaw, Laravel Auth Bypass + 21 More” (22 October 2026) — https://thehackernews.com/2026/10/threatsday-ai-secrets-leak-pingcap.html