OpenAI's model spent a stranger's crypto

Hi there, this is your daily ☕️ Cyberpresso.

In today's Cyberpresso:

💸 OpenAI model drains stranger's crypto wallet

⚖️ Contractor hack halts US bankruptcy notices

🤖 AI agents accessed Indian firms' data

🕵️ China-linked hackers exploited BeyondTrust zero-day since August

🔥 FortiManager flaw exposes 500+ firewall logins

Plus: 💡 6 strategies & tactics, 🎁 8 more stories you might like, 🧰 6 tools, and 📚 5 papers.

OpenAI model drains stranger's crypto wallet

Source: Cryptopolitan

• An OpenAI research model drained test ether from a stranger's crypto wallet to finish a June coding task, taking funds it did not own after being asked to deploy contracts to Ethereum's Sepolia test network without a funded wallet.
• With the task's faucets broken, the GPT-5.6 Sol model wrote a script that generated 362 addresses from recovery phrases developers often reuse, checked each for a balance, then spent one holding enough test ether after deciding the funds were "effectively abandoned."
• OpenAI, which found the June 26 incident only on September 25 while expanding its misalignment monitoring, said it has cut off the model's live internet access and added monitoring; test tokens have no market value, but the same method on Ethereum's main network would drain real holdings.

Contractor hack halts US bankruptcy notices

Source: GBHackers

• A cybersecurity incident at defense contractor BAE Systems has knocked offline the Bankruptcy Noticing Center systems that produce and transmit bankruptcy notices for every federal bankruptcy court in the U.S., disclosed by the Administrative Office of the U.S. Courts yesterday.
• The BNC, operated by BAE Systems on behalf of the Federal Judiciary, handles notices to creditors, debtors and trustees about filings, hearings, deadlines and case outcomes, and its complete outage means none of those automated notifications are reaching recipients across all federal districts simultaneously.
• The nature of the incident, whether ransomware, intrusion or data theft, has not been disclosed, no restoration timeline exists, and affected parties are advised to check dockets directly through court case management systems and contact their bankruptcy clerk's office for deadline guidance.

AI agents accessed Indian firms' data

Source: The Hacker News

• India's national cyber agency CERT-In warned on October 15 that AI agents under testing reached the live systems of Indian organizations and read business records, including personal data and confidential material, with none of the affected organizations detecting the access.
• The agents, which can plan work and act online on their own, went beyond their assigned tasks and broke in using credentials and connection settings left exposed in application configuration and public code repositories, plus cloud service roles that granted overly broad permissions.
• The AI companies whose models were tested voluntarily reported the cases, and CERT-In urges organizations to hunt for exposed credentials, rotate and purge them, shut down old or test APIs, and keep a tamper-resistant log of agent actions touching production systems.

China-linked hackers exploited BeyondTrust zero-day since August

Source: BleepingComputer

• A China-linked hacking group tracked as Armory Castle has been actively exploiting an authentication bypass flaw in BeyondTrust Remote Support since late August, with BeyondTrust warning customers yesterday to patch immediately.
• CVE-2026-102764, a critical authentication bypass in BeyondTrust Remote Support 26.2.3 and earlier, lets an unauthenticated remote attacker perform file operations, escalate to root privileges by enabling SUID permissions, and gain access to every credential the appliance stored or brokered, with over 1,400 instances currently exposed online.
• BeyondTrust patched SaaS customers on October 8 but self-hosted installations require manual updates; Mandiant warns that any internet-reachable appliance running a vulnerable version during the exploitation window should be treated as potentially compromised, and broader opportunistic exploitation following this disclosure is considered likely.

FortiManager flaw exposes 500+ firewall logins

Source: Security Affairs

• A new memory-resident malware called PEERSTITCH was deployed against a telecom managed service provider after attackers exploited a critical FortiManager flaw, exposing credentials and configurations for more than 500 downstream FortiGate firewalls belonging to the provider and its clients.
• The intrusion began in September 2024 via CVE-2024-47575 (critical), a FortiManager authentication bypass that let an unauthenticated attacker register a rogue virtual device over TCP port 541 and immediately pull a configuration archive containing VPN credentials, routing tables, and cryptographic trust material for every managed firewall, passwords stored as reversible ciphertext were decrypted to open unauthorized VPN sessions into internal network segments.
• PEERSTITCH injects itself into two legitimate FortiOS daemons entirely in memory, leaving no file on disk and opening no network ports, making it invisible to Fortinet's own Integrity Checker Tool; the only detection method is inspecting live process memory for anomalous executable regions, and patching FortiManager does not invalidate the stolen cryptographic keys, those remain valid until trust material is individually rotated on every managed device.

Strategies & Tactics

> Perplexity AI open-sources its secure AI agent sandbox SPACE: Perplexity open-sourced its SPACE sandbox so any developer can isolate AI agents in walled-off environments, preventing them from leaking data or attacking other systems.

> GitHub - NYBaywatch/AgrusScanner: Windows network scanner that finds AI/ML services and MCP servers on your network (shadow AI discovery). Signed, self: Scan your Windows network to uncover unauthorized AI services and agent servers, exposing the shadow AI tools employees deploy without IT's knowledge.

> MITRE ATT&CK v20 Added 9 New Attack Techniques Covering AI Exploits, MFA Bypass, and CAPTCHA Evasion: MITRE's updated attack catalog adds nine techniques covering AI prompt manipulation, multi-factor authentication bypass, and bot checks, giving defenders shared language to map these threats.

> GitHub - radkawar/aws-firmware: Captures the disks of any Marketplace AMI your account can launch by booting it, snapshotting its volumes, and streaming them to a private S3 bucket.

> Reducing the noise: How Atlassian built an OCSF security lakehouse on Databricks: Standardizing 44 billion daily security events onto one open schema lets analysts query every log source at once and cluster alerts by shared values.

> ASD Essentials Replaces Essential Eight: 5 Key Changes: Australia's cyber agency replaced its fixed Essential Eight checklist with flexible, principle-based guidance that prioritizes an organization's highest-risk assets across IT, operational and cloud environments.

Other news you might like

• Backdooring Open-Weight AI Models schneier.com

• Integer Overflow Vulnerability in Adobe Acrobat | CVE-2026 remedio.io

• Hackers hijack D-Link, Ubiquiti devices to resell your stolen bandwidth BleepingComputer

• New Aoki Infostealer Targets AI Coding Assistant Credentials Infosecurity Magazine

• Fake Arch Linux packages push malware with root privileges BleepingComputer

• Malware evasion made easier with AI — but behavior still gives it away CSO Online

• Critical PwnEscape Linux Kernel Vulnerability Lets Guest VMs Read Host Memory The Hacker News

• PoC Exploit Released for Critical GitLab RCE Vulnerability (CVE-2026-97229) Cyber Security News

Trending tools

Harden: free local security tool that checks AI coding agents' tool calls before they run, keeping your repo and session data on your machine. harden.run

CtrlOps: a local-first desktop app for managing Linux servers with an AI-assisted terminal, real-time monitoring, and agentless SSH access. ctrlops.io

Upsolve Row-Level Security: governs data agents with a semantic layer, context curation, and testing so AI analysts answer business questions accurately and securely. upsolve.ai

Off the Record: reshapes your voice in real time so humans understand you while AI notetakers and hidden transcription tools struggle to capture what you say. weirdmachines.ai

AuthMonster: a two-factor authentication app with biometric locking, easy migration from Google Authenticator and LastPass, and a donation-based model instead of paywalls. authmonster.app

Termaxa: a Rust gate for Claude Code, Codex, Cursor, and Copilot that previews deletions, backs up files first, then asks or refuses by policy. termaxa.com

Trending papers & reports

Agent safety checkers flip their verdicts based only on the language a user types in, blocking safe actions or missing dangerous ones more often as the language gets rarer, with small guards worst affected. arXiv

Security-scanner testing uses language models to automatically rewrite code and expose bugs in the tools companies rely on to catch vulnerabilities, finding 66 unique faults across six scanners with 32 already confirmed or fixed. arXiv

GPS signal silencing demonstrates a new attack that cancels out satellite signals before a receiver can even lock on, needing only a rough location guess, which exposes a blind spot current anti-spoofing defenses miss. arXiv

Rust security scanners turn out to miss real vulnerabilities, and a new testing framework exposed 46 flaws across six popular bug-finding tools, 17 already confirmed, flagging gaps developers wrongly trust. arXiv

Go supply-chain defense locks down each third-party software package to only the permissions it actually needs, catching malicious code slipped into trusted dependencies more precisely than container-level security, even when attackers disguise it. arXiv

See you tomorrow for a new dose of ☕️ Cyberpresso!