Where Secrets Really End Up in Docker Images: Testing 8 Common Methods
A comprehensive technical investigation examined exactly where secrets end up when building Docker images using eight different techniques. The tests used Ubuntu 24.04.1 LTS, Docker Engine 29.1.3, and BuildKit v0.26.2, with a test secret string SUPER_SECRET_KOTY_HULIGANY_2026 placed in various ways.
The first method, COPY, adds the secret file directly into an image layer. Even after running docker history, the file remains fully recoverable by extracting layers from a saved tar archive. Adding a subsequent RUN rm command only marks the file as deleted in a new layer; the original data stays intact in the previous layer.
Using ENV stores the secret value in the image configuration JSON under config.Env and also appears in build history. The ARG instruction keeps the value in history.created_by fields even though it does not appear in the final container environment.
The dedicated BuildKit Secrets feature, invoked with --mount=type=secret and --secret build flag, mounts the secret only temporarily during the RUN step. After the build completes, the value is absent from all layers, configuration JSON files, and history entries.
Multi-stage builds also proved effective when the secret is created and used only in an intermediate stage and never copied into the final image. No trace of the secret appears in the published artifact.
Combining echo and rm inside a single RUN instruction still leaves the secret string visible in build history and created_by metadata. Overwriting a copied file with a new value similarly fails because the original layer remains unchanged.
The experiments confirm that only two approaches — BuildKit Secrets and properly isolated multi-stage builds — avoid embedding secrets in the final Docker image. All other common patterns leave extractable data that can be recovered without ever starting a container.
Related articles
Security Researcher Builds SAST Scanner for AI-Generated Code and Audits 3,800 Public Repositories
A developer released AigisSAST, a lightweight open-source static analysis tool written in pure Python with no external dependencies, specifically tuned to detect common mistakes made by AI coding assistants. The scanner was run across roughly 3,800 repositories ranging from small pet projects to popular open-source platforms. It identified thousands of potential secrets and misconfigurations, yet manual review reduced the number of genuine leaks to approximately 30 cases, mostly Telegram bot tokens, database credentials, and committed .env files. The project also examined 471 production-grade Telegram bots handling payments and VPN services, uncovering 31 repositories that exposed real credentials either in current code or in Git history. AigisSAST includes 21 detection rules, 193 regression tests, automatic remediation via the fix command, and seamless integration with GitHub Actions. The author deliberately avoided validating any discovered keys to stay within ethical research boundaries.
Vendor Responsibility in Open Source: Licensing Obligations Exposed by Sonatype Nexus Changes
The article examines how vendors building products on copyleft open source projects like Nexus Repository OSS inherit significant legal and security responsibilities under licenses such as EPL 1.0. Sonatype's February 2025 shift from regular OSS binary releases to a limited Community Edition forces downstream vendors to handle their own builds, patch porting, and compliance disclosures. This change highlights the second part of copyleft licenses that outlines obligations for distributors, including revealing modifications and assuming liability for the final product. Security implications arise because critical vulnerabilities in the upstream project must now be tracked and patched by the vendor, with delays creating measurable supply chain risks. The piece provides a practical checklist for buyers to assess licensing hygiene, SBOM availability, and vulnerability response times in any open source-based solution.
PhantomSub Campaign Deploys 101 Malicious npm Packages to Hijack WhatsApp Accounts for Unauthorized Channel Subscriptions
Researchers at OX Security uncovered 101 malicious npm packages tied to the PhantomSub campaign that abuse connected WhatsApp accounts to subscribe users to promotional channels without consent. The packages disguise themselves as modified versions of the open-source Baileys library used for WhatsApp automation. Attackers rely on authenticated sessions rather than simple package installation, allowing them to control subscriptions through lists stored on GitHub, in plaintext, or as encoded identifiers. The packages have accumulated roughly 490,000 downloads, including 116,000 in the past 30 days, though the exact number of compromised accounts remains unknown. As of 28 September, npm had removed only 16 of the identified packages. The operation ultimately benefits channels selling bots, game resources, accounts, and promotion services by inflating subscriber counts while disabling notifications to hide the activity.
AI Model Hallucinations Fuel Slopsquatting Attacks on PyPI and npm Registries
Researchers identified 139 package names consistently hallucinated by five different AI models across Python and JavaScript ecosystems. Seven of these names are already registered on PyPI and npm, including one previously used to distribute malware. The attack vector, termed slopsquatting, allows attackers to register AI-suggested package names and execute code with developer privileges during installation. One package, metro-evaluator, contained malicious code removed by npm in December 2025, while another empty package css-color-stop began receiving downloads after the list was published. Real projects such as odf and lusid now occupy names that AI models recommend, causing developers to install unrelated software. Studies show hallucination rates between 4.62% and 21.7% depending on the model, with commercial models performing better than open-source ones. The findings highlight risks when AI coding agents execute dependency installation commands without human verification.