supabase

Orca Security secures AI-built and developer-created applications Orca Security has announced two new AI-powered capabilities: Orca AI AppGen Security, which discovers and secures AI applications built outside the development pipeline on AI-powered platforms like Claude, Supabase, and Lovable, and AI Code Security Auditor, which delivers deep AI-driven static analysis for code developed within traditional pipelines. The capabilities extend the Orca Platform to secure software across the full spectrum of application development, from professional engineering teams to the growing population of AI-assisted builders. Orca’s newly released State of AI Security Report 2026, based on anonymized telemetry from more than 1,200 production cloud environments analyzed by the Orca Research Pod, found that 52% of organizations now build custom applications with AI. As business teams increasingly deploy AI-generated applications connected to APIs, cloud resources, and sensitive data, software development is expanding beyond security’s traditional visibility. At the same time, exploitable risks continue to persist within development pipelines. IBM estimates breaches involving shadow AI cost organizations an average of $670,000 more than other incidents, underscoring the need for organizations to secure software consistently, whether it’s built by developers in the pipeline or employees using AI outside of it. “Everyone is a builder now. Developers are creating software in the pipeline, employees are building AI applications outside it, and the next generation of frontier AI models will increasingly generate and modify software on their own. Organizations need security that can keep pace with this shift without slowing innovation. Our AI Code Security Auditor prepares customers for the next wave of AI-assisted software development, while giving security teams the deep, accurate context they need to understand what’s truly exploitable. Combined with AI AppGen Security, Orca helps organizations secure every builder and every application, wherever and however it’s created,” said Gil Geron, CEO of Orca Security. Whether software is written by professional developers or generated by AI-powered builders, security teams need the same outcome: complete visibility, meaningful context, and the ability to prioritize real risk. Orca AI AppGen Security and Orca Code Security Auditor extend the Orca platform to secure software development wherever it happens. Orca AI AppGen Security: Secure the new generation of AI builders As AI expands software development beyond engineering, Orca AI AppGen Security gives security teams visibility and control over applications built outside the development pipeline by: Discovering AI-generated applications and who built them. Mapping application risk across APIs, integrations, and data access. Prioritizing high-risk exposures based on real business impact. Orca Code Security Auditor: Secure AI-assisted development Orca Code Security Auditor helps developers identify and prioritize the vulnerabilities that matter most by: Finding vulnerabilities traditional SAST misses with AI-powered code analysis. Full repository scanning to uncover Mythos class frontier model vulnerabilities. Prioritizing exploitable risk so teams can focus on what attackers can actually reach. “My developers and my business teams are both shipping software faster than my team can review it, and the apps built outside our pipeline were a complete blind spot. Seeing exploitable code and the AI apps our employees stand up in one platform lets us say yes to builders instead of slowing them down — and still know exactly what we’re governing,” said Sangram Dash, CISO at Sisense.

Secrets sprawl isn't slowing down: in 2025, it accelerated faster than most security teams anticipated. GitGuardian's State of Secrets Sprawl 2026 report analyzed billions of commits across public GitHub and uncovered 29 million new hardcoded secrets in 2025 alone, a 34% increase year over year and the largest single-year jump ever recorded. This year's findings reveal three core trends: AI has fundamentally reshaped how and where credentials leak, internal systems are far more exposed than most organizations realize, and remediation continues to be the industry's Achilles heel. Here are nine strategic takeaways that matter. 1. Secrets are growing faster than the developer population Since 2021, leaked secrets have grown 152%, while GitHub's public developer base expanded 98%. More developers and more AI-assisted code generation mean more credentials in circulation, and detection alone can't keep pace. 2. AI services drove 81% more leaks year over year GitGuardian detected 1,275,105 leaked secrets tied to AI services in 2025, up 81% from 2024. Eight of the ten fastest-growing categories of leaked secrets were AI-related. This isn't just about OpenAI or Anthropic keys. The real explosion is happening in LLM infrastructure: retrieval APIs like Brave Search (+1,255%), orchestration tools like Firecrawl (+796%), and managed backends like Supabase (+992%). Every new AI integration introduces another machine identity, and each one expands the attack surface. Deploying AI safely requires a proper secrets security strategy. 3. Internal repositories are 6x more likely to leak than public ones While public GitHub gets the attention, internal repositories are where the highest-value credentials live. GitGuardian's research found that 32.2% of internal repos contain at least one hardcoded secret, compared to just 5.6% of public repos. These aren't test keys. They're CI/CD tokens, cloud access credentials, and database passwords, the exact assets attackers target once they gain a foothold. Security through obscurity has failed. Treat internal repos as first-class leak sources. 4. 28% of leaks happen entirely outside code Secrets don't only live in repositories. GitGuardian found that 28% of incidents in 2025 originated entirely outside source code, in Slack, Jira, Confluence, and similar collaboration tools. These leaks are more dangerous: 56.7% of secrets found only in collaboration tools were rated critical, compared to 43.7% for code-only incidents. Teams share credentials during incident response, troubleshooting, and onboarding. If you're only scanning code, you're missing a quarter of your exposure. And the credentials leaking in collaboration tools are usually more critical and severe. 5. Self-hosted GitLab and Docker registries expose secrets at 3-4x the rate of public GitHub GitGuardian discovered thousands of unintentionally exposed self-hosted GitLab instances and Docker registries in 2025. Scanning these systems revealed 80,000 credentials, with 10,000 still valid. Secrets in Docker images were particularly troubling: 18% of scanned Docker images contained secrets, and 15% of those were valid, compared to 12% of GitLab repositories with a 12% validity rate. Docker secrets are also more production-adjacent. The perimeter between private and public is porous. 6. 64% of secrets leaked in 2022 remain valid today Detection is not remediation. GitGuardian retested secrets confirmed as valid in 2022 and found that 64% are still exploitable four years later. This is not a rounding error. It's proof that rotation and revocation are not routine, owned, or automated in most organizations. Credentials embedded across build systems, CI variables, container images, and vendor integrations are hard to replace without breaking production. For many teams, the safest short-term choice is to do nothing, leaving attackers with durable access paths. 7. Developer endpoints are the new credential aggregation layer The Shai-Hulud 2 supply chain attack gave researchers rare visibility into what secrets actually look like on compromised developer machines. Across 6,943 systems, GitGuardian identified 294,842 secret occurrences corresponding to 33,185 unique secrets. On average, each live secret appeared in eight different locations on the same machine, spread across .env files, shell history, IDE configs, cached tokens, and build artifacts. More striking: 59% of compromised machines were CI/CD runners, not personal laptops. Once secrets start sprawling into build infrastructure, they become an organizational exposure problem, not just an individual hygiene issue. More recently, the LiteLLM supply chain attack demonstrated the same pattern, with compromised packages harvesting SSH keys, cloud credentials, and API tokens from developer machines where AI development tools are increasingly concentrated. 8. MCP servers exposed 24,000+ secrets in their first year Model Context Protocol (MCP) made AI systems more useful by connecting them to tools and data sources. It also introduced a new class of credential exposure. In 2025, GitGuardian found 24,008 unique secrets in MCP-related config files on public GitHub, with 2,117 verified as valid. As agentic AI adoption accelerates, MCP and similar frameworks will normalize putting credentials into config files, startup flags, and local JSON. The agent ecosystem is expanding faster than security controls can adapt. 9. Shift from secrets detection to non-human identity governance The industry's limiting factor is answering three questions at scale: What non-human identities exist in my environment? Who owns them? What can they access? Organizations embracing agentic AI need to move beyond detection and build continuous NHI governance. That means eliminating long-lived static credentials wherever possible, adopting short-lived identity-driven access, implementing secrets vaulting as the default developer workflow, and treating every service account, CI job, and AI agent as a governed identity with lifecycle management. The Bottom Line Secrets sprawl is not slowing down. It's accelerating alongside AI adoption, developer productivity tools, and distributed software delivery. The old model of scanning public repos and hoping for compliance is no longer sufficient. Security teams need visibility across internal systems, collaboration tools, container registries, and developer endpoints. They need remediation workflows that can rotate credentials without breaking production. And most importantly, they need to stop treating secrets as isolated incidents and start managing them as part of a broader non-human identity governance program. The attack surface has changed. The question is whether security programs will change with it. About the Research GitGuardian's yearly State of Secrets Sprawl report was published for the 5th time, analyzing billions of public commits on GitHub, monitoring internal incidents across customer environments, and conducting original research on self-hosted infrastructure exposure and supply chain compromises.

The Pakistan-aligned threat actor known as Transparent Tribe has become the latest hacking group to embrace artificial intelligence (AI)-powered coding tools to strike targets with various implants. The activity is designed to produce a "high-volume, mediocre mass of implants" that are developed using lesser-known programming languages like Nim, Zig, and Crystal and rely on trusted services like Slack, Discord, Supabase, and Google Sheets to fly under the radar, according to new findings from Bitdefender. "Rather than a breakthrough in technical sophistication, we are seeing a transition toward AI-assisted malware industrialization that allows the actor to flood target environments with disposable, polyglot binaries," security researchers Radu Tudorica, Adrian Schipor, Victor Vrabie, Marius Baciu, and Martin Zugec said in a technical breakdown of the campaign. The transition towards vibe-coded malware, aka vibeware, as a means to complicate detection has been characterized by the Romanian cybersecurity vendor as Distributed Denial of Detection (DDoD). In this approach, the idea is not to sidestep detection efforts through technical sophistication, but rather to flood target environments with disposable binaries, each using a different language and communication protocol. Helping threat actors in this aspect are large language models (LLMs), which lower the barrier to cybercrime and collapse the expertise gap by enabling them to generate functional code in unfamiliar languages, either from scratch or by porting the core business logic from more common ones. The latest set of attacks has been found to target the Indian government and its embassies in multiple foreign countries, with APT36 using LinkedIn to identify high-value targets. The attacks have also singled out the Afghan government and several private businesses, albeit to a lesser extent. The infection chains likely begin with phishing emails bearing Windows shortcuts (LNKs) bundled within ZIP archives or ISO images. Alternatively, PDF lures featuring a prominent "Download Document" button are used to redirect users to an attacker-controlled website that triggers the download of the same ZIP archives. Regardless of the method used, the LNK file is used to execute PowerShell scripts in memory, which then download and run the main backdoor and facilitate post-compromise actions. These include the deployment of known adversary simulation tools like Cobalt Strike and Havoc, indicating a hybrid approach to ensure resilience. Some of the other tools observed as part of the attacks are listed below - Warcode, a custom shellcode loader written in Crystal that's used to reflectively load a Havoc agent directly into memory. NimShellcodeLoader, an experimental counterpart to Warcode that's used to deploy a Cobalt Strike beacon embedded into it. CreepDropper, a .NET malware that's used to deliver and install additional payloads, including SHEETCREEP, a Go-based infostealer that uses Microsoft Graph API for C2, and MAILCREEP, a C#-based backdoor utilizing Google Sheets for C2. Both malware families were detailed by Zscaler ThreatLabz in January 2026. SupaServ, a Rust-based backdoor that establishes a primary communication channel via the Supabase platform, with Firebase acting as a fallback. It contains Unicode emojis, suggesting that it was likely developed using AI. LuminousStealer, a likely vibe-coded, Rust-based infostealer that uses Firebase and Google Drive to exfiltrate files matching certain extensions (.txt, .docx, .pdf, .png, .jpg, .xlsx, .pptx, .zip, .rar, .doc, and .xls). CrystalShell, a backdoor written in Crystal that's capable of targeting Windows, Linux, and macOS systems, and uses hard-coded Discord channel IDs for C2. It supports the ability to run commands and gather host information. One variant of the malware has been found to use Slack for C2. ZigShell, a counterpart to CrystalShell that's written in Zig and uses Slack as its primary C2 infrastructure. It also supports added functionality to upload and download files. CrystalFile, a simple command interpreter written in Crystal that continuously monitors the "C:\Users\Public\AccountPictures\input.txt" and executes the contents using "cmd.exe." LuminousCookies, a Rust-based specialized injector to exfiltrate cookies, passwords, and payment information from Chromium-based browsers by circumventing app-bound encryption. BackupSpy, a Rust-based utility designed to monitor the local file system and external media for high-value data. ZigLoader, a specialized loader written in Zig that decrypts and executes arbitrary shellcode in memory. Gate Sentinel Beacon, a customized version of the open-source GateSentinel C2 framework project. "The transition of APT36 toward vibeware represents a technical regression," Bitdefender said. "While AI-assisted development increases sample volume, the resulting tools are often unstable and riddled with logical errors. The actor's strategy incorrectly targets signature-based detection, which has long been superseded by modern endpoint security." Bitdefender haș warned that the threat posed by AI-assisted malware is the industrialization of the attacks, allowing threat actors to scale their activities quickly and with less effort. "We are seeing a convergence of two trends that have been developing for some time: the adoption of exotic, niche programming languages, and the abuse of trusted services to hide in legitimate network traffic," the researchers said. "This combination allows even mediocre code to achieve high operational success by simply overwhelming standard defensive telemetry."