An investigation into an intercepted SSH session uncovered 333 hijacked Tor exit relays directing connections to password-stealing servers. Separately, an analysis of Heartbreaker ransomware found a macOS sample that carries its own source code and attempts to rebuild itself after encrypting files.
The findings feature in SecurityWeek’s cybersecurity roundup, which also covers stolen developer-platform credentials, Hookedge backdoor intrusions and banks’ restrictions on third-party AI agents.
Password interception through compromised Tor infrastructure
The Tor investigation began when a researcher’s own SSH connection was intercepted. It revealed that roughly a tenth of all exit relays—333 in total—were sending SSH traffic to 12 servers controlled by attackers. Users who dismissed SSH’s warning and entered their passwords exposed those credentials. The attackers subsequently tried to install Ebury on the servers the users had intended to reach.
Ebury had already compromised the machines hosting the malicious relays. Changes to their exit policies also prevented connections to encrypted email retrieval ports while allowing connections to plaintext ports.
Within three days, the Tor Project removed the relays. One affected operator identified password reuse as the cause of its own compromise. SecurityWeek said the project had not answered its request for comment.
Heartbreaker’s rebuilding mechanism has practical limits
Patrick Wardle examined Heartbreaker, a ransomware sample written in Rust for macOS and submitted to VirusTotal in May 2025. No antivirus engine on the service detected it as of his writing.
Its execution sequence begins with an attempt to create a LaunchAgent for persistence, followed by encryption of files in the user’s Downloads folder. The sample then attempts to produce a new hash through recompilation: it alters identifiers in its embedded source code, introduces junk functions and builds another copy. If Rust is absent, it even attempts to install it.
Wardle considers the current sample unlikely to become a major threat. A single hardcoded AES key allows files to be recovered for free, while the rebuilding mechanism requires a Rust framework that is no longer publicly available.
Railway pauses deployments as PingCAP details stolen records
Railway has halted all deployments and called on every customer to rotate stored environment variables and private registry credentials. The developer platform says an unauthorized party may have accessed encrypted customer environment variables and recovered some of their contents in decrypted form. Such variables can contain API keys, database credentials and registry credentials.
The company believes the incident forms part of a series of attacks against platform companies targeting AI inference providers. It has notified those providers to revoke keys proactively. Railway says the affected population may be limited to a subset of users, despite its platform-wide rotation request.
At PingCAP, stolen service account credentials enabled an unauthorized party to enter some internal systems between September 9 and October 7, according to the TiDB developer. The intruder copied data and obtained source code.
The exported information covered TiDB Cloud user names and email addresses, customer contacts, contract and billing details, and encrypted values of TiDB Cloud API keys. PingCAP says it has found no evidence of an impact on data stored in TiDB. It has invalidated every identified compromised credential and remedied four vulnerabilities reported by an external firm. Work to identify affected customers continues.
Bullhorn is also restoring service after a cyber incident led it to disable access to its systems. Whether customer data was exfiltrated remains undetermined. During the restoration of its ATS/CRM platform, users without SSO must reset passwords, and API users must rotate their API passwords. The company reports no indication of an impact on Bullhorn Automation, Sourcebreaker or VMS Sync.
Symantec attributes Hookedge intrusions to APT28
A naval shipbuilder, a military aircraft engine manufacturer and a law enforcement agency in two NATO member states were targeted with the Hookedge backdoor, according to Symantec. The company attributes the attacks to Russia’s APT28, which it calls Swallowtail.
The shipbuilder intrusion persisted for at least two months. There, the attackers used two credential-theft techniques and maintained an SSH tunnel into the internal network. They also deployed a loader that Symantec had not previously documented; it decrypts its payload and executes it in memory.
Banks confront AI activity inside customer accounts
Axios reports, citing sources, that major US banks have detected unauthorized third-party AI agents operating in customer accounts. Some attempted actions appear to have occurred independently of user instructions, although many banks cannot determine whether customers directed the activity.
Customers typically supply the agents with permission and login credentials, contrary to most banks’ rules. Morgan Stanley has started blocking third-party agents, and JPMorgan, Bank of America and Citi are restricting them as well. Most banks are moving toward read-only access.
Separately, ISO and IEC have issued ISO/IEC 27090 to guide detection and mitigation of AI-specific threats throughout systems’ life cycles, including data poisoning and model theft. It complements ISO/IEC 27001 and 27002 and highlights how retraining could carry corrupted operational data into later model versions.
Investigators examine ShinyHunters contacts; courts weigh evidence and compensation
KrebsOnSecurity, citing sources close to a federal investigation, reports that investigators believe at least one executive at a US-based consulting firm helped run Shinywire. The now-defunct ShinyHunters site was used to pressure nearly 100 hacked organizations to pay.
Authorities accessed Signal messages when they arrested ShinyHunters suspect Pepijn van der Stap in the Netherlands. Those messages reportedly show frequent contact with Canadian cybersecurity executive Edward Dubrovsky and at least one executive at a US consulting firm, with contacts sometimes aware of new victims before the victims themselves.
In litigation involving WhatsApp, a federal judge declined to find NSO Group in contempt over allegations that it breached a 2025 injunction prohibiting use of WhatsApp’s infrastructure to target users. The sole evidence was a Meta investigator’s declaration identifying 23 new accounts with signatures previously associated with NSO testing. The judge found the evidence fell short of the clear-and-convincing standard, noting that WhatsApp had withheld detail to protect its detection methods.
An $80 million proposed settlement in a separate case would resolve a consolidated class action over Conduent’s 2024 breach, which involved theft of millions of people’s personal and health information. Subject to court approval, all class members could claim a year of credit or medical data monitoring and reimbursement of documented losses up to $3,500. A pro rata cash payment would also be available to people whose stolen data was in a readily accessible format. Conduent disputes the allegations of wrongdoing and liability.
