Daily brief at 7am Melbourne. Unsubscribe any time.

Wednesday 20 May 2026

The Cobbler's Children: CISA Left Its Own Credentials in a Public GitHub Repo for Six Months

CISA — America's cyber defence agency — left SSH keys, plaintext passwords, and AWS credentials in a public GitHub repo for six months, and Congress wants answers.

Lead story

The Cobbler's Children: CISA Left Its Own Credentials in a Public GitHub Repo for Six Months

America's cybersecurity watchdog — the agency that publishes advisories telling everyone else to rotate credentials, segment networks, and assume breach — spent six months with SSH keys, plaintext passwords, AWS workspace credentials, and authentication tokens sitting in a public GitHub repository. The repo was called, with exquisite irony, "Private-CISA."

A researcher who stumbled across it described it as one of the worst credential exposures they'd ever seen. The filenames weren't even subtle: external-secret-repo-creds.yaml and AWS-Workspace-Firefox-Passwords.csv were reportedly among the files sitting there since November 2025. By the time it was flagged and taken down, it had been publicly accessible for roughly six months.

Why this stings more than the average credential leak

CISA is the agency that runs "Secure by Design" campaigns, publishes joint advisories with the Five Eyes partners (including the ASD here in Australia), and lectures critical infrastructure operators about third-party risk. The reputational damage isn't just embarrassing — it hands adversaries a rhetorical crowbar to prise open any future CISA guidance with "why should we listen to you?"

The practical risk depends on what exactly was in those credentials and whether they've been rotated since, which CISA hasn't confirmed publicly. But six months is a long time. Automated credential-harvesting bots scan GitHub continuously — it's a known attack vector with its own tooling. If any of those keys were live for that window, the question isn't whether someone found them; it's whether anyone used them.

Congress is now involved

Capitol Hill has demanded answers, with legislators asking CISA to explain how the exposure happened, when it was discovered, what access the credentials permitted, and whether any systems were subsequently compromised. That inquiry has real teeth given CISA's already-fraught political environment: the agency has faced staffing cuts and leadership uncertainty over the past year.

What it signals about secrets management at scale

The irony is that this is a solved problem. Tools like HashiCorp Vault, AWS Secrets Manager, and GitHub's own secret scanning (which CISA presumably knows about) exist precisely to catch this. The failure here isn't exotic — it's the same mistake that causes breaches at companies CISA warns every week. Someone committed secrets to a repository, the repo was either accidentally made public or was always misconfigured as public, and no automated scanning caught it.

For Australian organisations, the ASD's Essential Eight and the Protective Security Policy Framework both mandate credential management controls that would, in theory, prevent exactly this. The lesson isn't schadenfreude — it's that even the teams who write the playbooks can get caught skipping steps.

What to watch

CISA's response to the Congressional inquiry will tell us a lot. If the agency can demonstrate the credentials were rotated promptly and there's no evidence of downstream compromise, the story ends as an embarrassing misconfiguration. If those keys touched production systems that weren't immediately locked down, it becomes something considerably worse. Either way, expect this to fuel renewed pressure on CISA to demonstrate it can manage its own house before managing everyone else's.

Also today

Google I/O 2026: Agents All the Way Down

Google's annual developer conference landed like a full-stack rebrand. The headline product is Gemini Spark, an always-on agentic assistant wired into Gmail, Calendar, and Search. Gemini 3.5 Flash — faster, cheaper, and built for autonomous task execution — underpins most of it. Google also redesigned the 25-year-old search box into a dynamic, AI-driven interface, and DeepMind CEO Demis Hassabis closed the keynote by saying we're standing in the 'foothills of the singularity.' Whether that's insight or hype, the product direction is clear: Google is building for agents, not chatbots. Australian Workspace and Google Cloud customers will be among the first in the Asia-Pacific region to access several of these features.

TechCrunch

Microsoft Kills a Malware-Signing-as-a-Service Racket

Microsoft has dismantled Fox Tempest, a cybercrime operation that hijacked the company's own Artifact Signing service to issue fraudulent code-signing certificates to ransomware gangs and other malware distributors. The certificates made malicious software appear legitimate to Windows Defender and endpoint tools, essentially laundering malware through Microsoft's own trust infrastructure. The company unsealed a US District Court filing detailing the disruption and confirmed its own machines were among the roughly thousands of victims. Fox Tempest had been operating since at least May 2025. This is a rare case of a platform being weaponised against itself — and a reminder that code-signing trust is only as strong as the verification behind it.

The Record

ChromaDB Has an Unpatched RCE — and It's Aimed at AI Apps

A maximum-severity vulnerability in ChromaDB, a popular open-source vector database used in AI and LLM applications, allows unauthenticated remote attackers to execute arbitrary code on any exposed server. The flaw lives in the Python FastAPI implementation and has no patch available yet. ChromaDB is widely used as the memory layer for AI agent pipelines — it's the thing that stores and retrieves embeddings for chatbots, RAG systems, and agentic workflows. An internet-exposed ChromaDB instance is essentially an unlocked door into the AI backend. If you're running ChromaDB in any environment reachable from the internet, take it offline or firewall it now.

Bleeping Computer

Shai-Hulud Returns: 600 npm Packages Poisoned in Supply-Chain Wave

Threat actors pushed more than 600 malicious packages to the npm registry as part of a renewed Shai-Hulud supply-chain campaign, with a related wave also compromising the @antv ecosystem — including echarts-for-react, a React wrapper with roughly 1.1 million weekly downloads. The attack compromised maintainer accounts to steal publishing tokens, install OS-level backdoors, and persist inside CI/CD pipelines. A separate GitHub Actions vector was also used, with tags on the actions-cool/issues-helper workflow redirected to imposter commits to harvest CI credentials. Taken together, these are not isolated incidents — they're a coordinated campaign against the open-source software supply chain. Any developer pulling from npm or relying on affected GitHub Actions workflows should audit their recent build logs.

CyberScoop

OpenAI and Google Unite on AI Content Watermarking

OpenAI announced it is adopting Google's SynthID watermarking technology and joining the C2PA open standard to embed verifiable provenance data into AI-generated images. Google separately confirmed that OpenAI, Nvidia, and others are now implementing SynthID across their platforms. The moves represent a rare moment of cross-industry alignment on a genuinely thorny problem: as AI-generated imagery becomes indistinguishable from photographs, the only scalable way to label it is to bake the label into the content itself rather than relying on platform moderation. Australia's eSafety Commissioner has flagged synthetic media as a priority concern, and this kind of provenance infrastructure is likely to inform any future Online Safety Act obligations around deepfakes.

OpenAI Blog

Australian Universities Hit by Canvas LMS Cyberattack

Multiple Australian universities have been caught up in a cyberattack targeting Instructure's Canvas learning management system, in what The Mandarin describes as a textbook case of platform concentration risk. Canvas dominates the Australian higher-education LMS market, meaning a single successful attack on the vendor has cascading effects across dozens of institutions simultaneously. The incident echoes concerns raised under Australia's Security of Critical Infrastructure Act about third-party and supply-chain risk — universities fall under the SOCI framework's higher education sector designation. Details of the specific vector and scope remain limited, but the pattern is familiar: one widely-deployed vendor, many dependent institutions, one point of failure.

The Mandarin

DirtyDecrypt PoC Drops for Linux Kernel Privilege Escalation Bug

A proof-of-concept exploit has been published for CVE-2026-31635, a local privilege escalation vulnerability in the Linux kernel nicknamed DirtyDecrypt (also DirtyCBC). The flaw was patched in April after being discovered by the Zellic and V12 security teams, though the maintainers noted it was a duplicate of an earlier-reported bug. The PoC's release dramatically shortens the window for unpatched systems — any Linux environment where a low-privileged user can run code is now at elevated risk. Cloud workloads, container hosts, and shared Linux infrastructure should be prioritised for patching. The Linux kernel runs the vast majority of Australia's cloud and government server infrastructure.

SecurityWeek

Huawei Zero-Day Was Behind Luxembourg's Nationwide Telecoms Outage

A Huawei zero-day vulnerability was responsible for last year's collapse of Luxembourg's entire national telecommunications network — a fact that has only now been disclosed publicly. The flaw remains unexplained, has not been assigned a CVE, and Huawei has not publicly acknowledged it. The Luxembourg outage at the time was one of the most significant single-country telecom failures in recent European history. The revelation raises obvious questions about similar Huawei infrastructure deployments elsewhere. Australia's position on Huawei in 5G networks is well-established, but the vendor's equipment remains present in older network layers in various markets globally, and an unacknowledged zero-day with no public patch is precisely the kind of finding that keeps network security teams awake.

The Record

Drupal Drops Urgent Patch for Highly Critical Core Vulnerability

Drupal's security team issued an emergency patch for all supported branches of its CMS on Wednesday, warning that exploits could emerge within hours of the release. The team deliberately withheld details of the vulnerability ahead of the patch to limit the exploitation window. Drupal powers a significant slice of government and enterprise websites globally — Australia's federal and state governments both run Drupal-based public-facing sites, making this a priority update for web teams. The patch also extends back to the unsupported 8.9 branch, which suggests the maintainers consider the vulnerability serious enough to warrant reaching back to legacy installs. If you're running Drupal, the patch timeline here is measured in hours, not days.

SecurityWeek

Discord Turns On End-to-End Encryption for All Voice and Video Calls

Discord has enabled end-to-end encryption by default for all voice and video calls across its platform, covering its hundreds of millions of users globally. The move means Discord itself can no longer access the content of calls — a meaningful shift for a platform that has historically operated with more visibility into user communications than Signal-style apps. The encryption uses a standard protocol and comes with a verification mechanism so users can confirm they're talking to who they think they are. For Australian users — including the large gaming and community server population — this is a genuine privacy upgrade that arrives without any action required on their end.

Bleeping Computer

OpenAI Launches Formal AI Partnership with Singapore Government

OpenAI has announced a multi-year partnership with Singapore to expand AI deployment across business and public services, build local technical talent, and support the city-state's AI governance ambitions. The deal is framed as a model for national AI partnerships in the Asia-Pacific region. It's a notable move in Australia's neighbourhood: Singapore has been aggressively positioning itself as the region's AI hub, and OpenAI's formal government partnership there creates a reference point that Australian policymakers and enterprise buyers will inevitably be asked about. Australia's own AI policy framework — currently under development following the Albanese government's AI regulation consultation — does not yet include anything equivalent to a formal national model deployment agreement.

OpenAI Blog

Sources consulted