AI
3 articles with this tag
Japan's Children and Families Agency Will Not Operate the 'Komamorou' Record-Check System for Now: Many Serious Vulnerabilities Found, and What Schools, Childcare and After-School Operators Should Check (Not a Data Leak)
On November 18, 2026, Japan's Children and Families Agency announced that the 'Komamorou system', developed so that operators and staff can check past sex-offence records (criminal-record checks) under the child sexual-violence prevention law, which is due to take effect on December 25, will not be operated for the time being. After the Prime Minister instructed on October 26 that a vulnerability check using high-performance AI be carried out ahead of schedule, a vulnerability assessment of the system found many serious vulnerabilities, and the agency says handling them before the effective date became difficult. The minister apologised to those concerned and told officials to urgently consider alternative procedures that do not use the system. The agency plans to compile alternatives around early December. This is about vulnerabilities in a system that had not yet gone into operation; the agency did not announce that personal data leaked. This site's take: affected operators should stop preparing on the assumption that the system will be used, wait for the alternatives due in early December, and check the agency's official page.
LiteLLM 1.107.0 Shipped a Credential Stealer (October 2026): Affected PyPI Release and Docker Images, and What Users Should Do
On October 21, 2026 (UTC), BerriAI, the company behind LiteLLM, announced that its release pipeline had been compromised and that a v1.107.0 containing malware that steals credentials and loads remote-code-execution tools had been published. Affected were litellm==1.107.0 on PyPI (11:39 to 13:23 UTC) and the v1.107.0, latest, main-latest and main-stable tags on Docker Hub (11:46 to 14:20 UTC) and GHCR (from 11:46 UTC until taken down by 13:51 UTC, about two hours). The official v1.107.0 source tag on GitHub does not contain the malicious files; the artifacts were built from an attacker-controlled lookalike repository. All earlier versions are outside this compromise, and the company asks users to pin 1.106.2 or another explicit earlier version. On October 24 (UTC) the company published 1.103.8, 1.104.6, 1.105.4 and 1.106.3, and from these releases Docker images and Helm charts are signed with a new cosign key (the previous key no longer verifies them). On October 23 the company published a root cause analysis and corrected its first belief that the tokens were taken from the logs of an internal LiteLLM deployment. According to the analysis, a configuration file committed by mistake to a public repository on May 27 contained the master key of an internal agent platform; the commit was overwritten but the key was never revoked. On October 10 the attacker used that key to call the platform's API and obtained the GitHub CLI token of a GitHub organization owner configured there. With it, the attacker exfiltrated CI secrets to obtain two more maintainers' tokens, and on October 21 disabled branch protection, changed the release workflow and approved the release. This site's conclusion: rebuild any environment that installed or started 1.107.0 during the window, and rotate every key it could reach, including LLM provider keys, the proxy master key, the database password and cloud roles.
What is shadow AI? The risks of AI use your company can't see, and what organizations and individuals should do
Shadow AI is employees using AI tools the company hasn't approved or doesn't manage (chatbots, AI browser extensions, coding assistants, meeting-note AI, AI agents connected to work accounts) for work. It's the AI form of shadow IT. The main risks: confidential or personal data entered may be stored by the provider and, depending on plan and settings, used for training (consumer and business terms differ); OAuth grants and API keys hand third-party tools access to mail and drives; output errors and licensing problems; and privacy-law and audit obligations. A ban alone pushes usage out of sight, so provide an approved tool on business terms that don't train on inputs, a short rule on what data may be entered, SSO and admin controls, regular review of OAuth grants and extensions, and an ask-first channel.