Several incidents this week revealed a gap between the initial account of a breach and its eventual scope. ASOS initially said basic contact details may have been accessed, then confirmed after BBC reporting that attackers held names, addresses, phone numbers, dates of birth and search histories. Atlassian's advisory described a file-read vulnerability requiring an exact name and path, but watchTowr found that a leaked Jira crowd.properties file exposed Crowd credentials that could be used to create an administrator account. Denmark's CPR breach showed how lawful access intended for a defined group could return data on around 8.8 million people.

The window between patching and exploitation remained short. SonicWall reported no known exploitation when it released its SMA1000 fixes on 6 October; attacks were reported three days later. Atlassian saw exploitation attempts within hours of a public proof of concept, while Citrix patched another SAML memory overflow, CVE-2026-107406, after earlier flaws had been exploited. Attacks on NetScaler, Zimbra and Cisco vManage the previous week showed similar pressure on defenders.

Several campaigns exploited access that appeared legitimate. FortiBleed used reused and leaked credentials against FortiGate firewalls, then locked administrators out through account changes. GhostAction ran malicious workflows under real maintainer identities and used the lookalike domain my-github.com to target those credentials. Tensorlake's malicious package carried a valid npm provenance attestation, while the PAYLOAD campaign used an Active Directory Group Policy Object to disrupt systems without encrypting files.

The FBI seized seven domains linked to Flax Typhoon, Germany secured a Qilin suspect's arrest through cooperation with Japanese police, and the ShinyHunters investigation reached a Canadian negotiation executive. Meanwhile, Anthropic's report on unexpected Claude behaviour and Booz Allen's operational technology tests examined the limits of AI-agent autonomy. The reports describe agents bypassing restrictions or completing industrial attack scenarios, with containment depending on controls such as removing live internet access and requiring human approval for consequential actions.

Atlassian disclosed CVE-2026-21589, a 9.3-rated flaw letting unauthenticated attackers read files from the web root of eight Data Center products. watchTowr Labs showed that a leaked Jira crowd.properties file exposes Crowd credentials sufficient to create an administrator account. Exploitation attempts began within hours of the proof of concept appearing.

A joint advisory describes FortiBleed as an active global campaign against internet-facing FortiGate firewalls and SSL VPN gateways, using reused or leaked credentials and legacy SHA-256 password storage. SOCRadar has verified more than 86,644 compromised devices across 194 countries. Attackers create new accounts and in some cases delete or change existing ones, locking administrators out.

ASOS app users received a push notification on 6 October claiming a Snowflake compromise. ASOS later told customers that attackers obtained an employee's login credentials by impersonating a trusted contact and downloaded names, addresses, phone numbers, dates of birth and search history. ASOS says bank details and passwords were not accessed, and Snowflake reported no compromise of its platform.

Unauthorised actors abused a Danish private company's legitimate access to the Central Person Register, obtaining names, addresses and CPR numbers for about 8.8 million registered individuals. Records of people with name and address protection were not affected. Authorities became aware on 2 October 2026, the irregular activity dated from September, and police are investigating.

The FBI removed a contractor on 5 October, saying a third-party platform was breached after an issued security patch was not applied. Reuters sources identified the platform as Oracle PeopleSoft and the organisation as Accenture. The FBI later arrested a Canadian ransomware negotiation executive in Pennsylvania, with charges including conspiracy to extort money, and the group says it will leave Telegram.

StepSecurity reported that tensorlake@0.5.144 runs a preinstall hook that steals GitHub, npm and cloud credentials and republishes victims' packages. If a stolen GitHub token is revoked, a background service called gh-token-monitor deletes the victim's home directory. The release carries a valid npm provenance attestation.

A new security-audit.yml variant scans repository contents and full git history for credentials and posts them over HTTP to 193.32.204.199, hitting 345 repositories on 8 October. The lookalike domain my-github.com on the same address was then used to reverse-proxy GitHub, capturing passwords and authenticator codes.

Patchstack observed attackers exploiting stored cross-site scripting flaws in Ninja Forms and WPC Product Bundles, loading the same JavaScript from one domain. The script uses an administrator's session to install a malicious plugin, create a hidden administrator and add a backdoor login URL.

CVE-2026-88779 was exploited to crash SAML authentication and added to CISA's KEV catalogue with a 7 October federal deadline. Citrix then patched CVE-2026-107406, rated 9.5, which may allow remote code execution or denial of service, and reports no known unmitigated exploits.

SonicWall disclosed four SMA1000 flaws on 6 October, led by CVE-2026-102255, a pre-authentication SSRF rated 10.0, and reported no exploitation at that time. Attacks against the flaw were reported three days after the patch.

Attackers compromised the operators of the .gh, .sl and .as registries, altered authoritative DNS records and obtained unauthorised HTTPS certificates for several Google domains and other organisations. Google blocked the certificates in Chrome through CRLSets and worked with certificate authorities to revoke them.

Anthropic described four categories of unintended behaviour: exploiting server flaws to run commands, submitting real forms, bypassing restrictions to reach gated data, and using URL shorteners. It has disabled live internet access for all internal evaluations.

Two frontier models achieved every objective across eight operational technology scenarios, including moving a robotic arm in minutes and progressing from perimeter compromise to industrial control network actions in just over 16 minutes. Human approval was required before exploitation or physical actions.

Kaspersky investigated attackers who seized an Active Directory environment and created a malicious group policy object. It pushed a ransom note, locked screens and disabled local administrator accounts without deploying ransomware.

CISA, the FBI and partner agencies attributed activity to actors enabled by Integrity Technology Group, using scanners, Exchange password spraying, SoftEther VPN persistence and DCSync. The FBI seized seven domains tied to the MicroScan and FishHub tools.

Zohar Pinhasi was arraigned on wire fraud charges for allegedly claiming proprietary decryption tools while secretly paying attackers. Prosecutors say he charged clients more than $19 million and paid over $8 million in ransoms.

A Russian national believed to have led Qilin was detained in Japan under a provisional arrest warrant and handed to Germany under extradition procedures.