AI bug bounty boom: from 'slop' to security game-changer

decrypt ↗

The gist

AI-powered bug bounty programs have rocketed from chaotic 'slop' to game-changing security engines—but the human touch is still mission-critical to avoid a new breed of automated risk.

What to know

AI's Growing Pains

AI bug hunters dazzled with deep finds but swamped maintainers with fabricated evidence and 'AI slop,' forcing a reckoning before models finally matured into trusted security partners.

The early adoption of AI in security research showcased promising capabilities, such as autonomously uncovering multiple vulnerabilities—including seven in FFmpeg and a critical IDOR missed by human testers—demonstrating AI’s potential to augment traditional methods. However, this promise was tempered by the significant challenge of validating AI-generated findings, as AI often fabricated validation evidence, turning the verification process into a resource-intensive endeavor that sometimes consumed more time than manual testing itself.

The surge of AI-generated bug reports overwhelmed bug bounty programs, with projects like Curl shutting down their initiatives due to an influx of low-quality, unvalidated submissions dubbed 'AI slop.' This flood was likened to a denial of service attack on these programs, eroding trust and forcing maintainers to expend considerable effort triaging and filtering out bogus reports, which in turn shifted the workload rather than reducing it.

Despite these early hurdles, advancements by early 2026—exemplified by AI models like Anthropic’s Claude Opus 4.6—marked a turning point where AI tools could reliably find meaningful zero-day vulnerabilities in well-tested codebases without specialized scaffolding. This evolution improved the quality of AI-assisted reports significantly, enabling maintainers like Daniel Stenberg to finally trust and integrate AI findings into their workflows, signaling a maturation from the 'completely useless' AI of 2024 to a 'very useful' security asset.

The rise of AI-driven vulnerability discovery is reshaping bug bounty paradigms, as companies increasingly consider leveraging internal AI scanners to preemptively identify flaws, potentially reducing reliance on external bounties and altering payout dynamics. Nonetheless, human expertise remains indispensable—not only for validating AI findings but also for the nuanced communication and persuasion involved in penetration testing, underscoring that AI, while powerful, cannot replace the critical human element in security assessment.

Sources
Paul's Security Weekly (Video)Application Security Weekly (Video)Security Weekly - A CRA ResourceThe PrimeTimeSecurity Weekly - A CRA Resource

LLMs Redefine Code Audits

Large language models have shifted from noisy static analyzers to creative, context-aware security agents—yet their outputs still demand careful human review before action.

By early 2026, large language models (LLMs) began transforming secure code analysis by efficiently parsing vast codebases to prioritize vulnerabilities and reduce manual review times, as seen with tools like CodeQL which initially flagged thousands of issues requiring extensive human triage. Beyond noise reduction, LLMs improved the accuracy of static analysis queries, aiming to cut false positives at the source, while also serving as creative brainstorming partners that prompt security researchers to explore overlooked attack vectors, enhancing both efficiency and insight generation.

Anthropic’s Claude platform exemplifies the leap toward AI-embedded security scanning, integrating vulnerability detection and patch suggestion directly into the coding environment after extensive stress testing. Claude Code Security’s reasoning-based approach, which contextualizes data flows and component interactions, uncovered over 500 previously unknown high-severity vulnerabilities, including subtle heap buffer overflows missed by traditional scanners. Despite these advances, human oversight remains essential, as findings require severity assessment and approval before remediation, underscoring the current gap between research prototypes and production-grade application security.

In a competitive response to Anthropic, OpenAI launched Codex Security in research preview, marking a strategic shift from commoditized AI models to specialized application-layer security agents. Codex Security leverages tailored prompts and tooling to proactively identify vulnerabilities across massive codebases, surfacing thousands of critical issues with improved contextual awareness. This evolution reflects the dual-edged nature of AI in security: while AI-generated code inflates the attack surface and introduces new risks, AI agents simultaneously accelerate vulnerability detection and remediation, albeit with the caveat that multiple runs and human review are necessary due to the non-deterministic nature of LLM outputs.

GitHub’s Dependabot automation, launched in early 2026, operationalizes AI-driven remediation by assigning vulnerabilities to AI coding agents that analyze specific codebases and propose fixes via draft pull requests for human engineers to review. This human-in-the-loop model dramatically shortens remediation times from industry averages of 8–70 days to mere hours, addressing the surge in security alerts fueled by the proliferation of AI-generated code and its novel dependency patterns. As GitHub’s team monitors over 20 million repositories across 32 ecosystems, this scalable integration of AI into security workflows exemplifies the practical balance between automation and necessary human judgment.

Sources
Security Weekly - A CRA ResourceSecurity Weekly - A CRA ResourceCybersecurity HeadlinesResilient CyberCISO Talk by James AzarMixture of Experts

Bug Bounty Survival Tactics

Open source and commercial bug bounty programs weathered an AI-fueled storm by evolving triage strategies, adopting phased launches, and embracing AI as both threat and indispensable ally.

In 2024, a landmark collaboration among tech giants including Anthropic, AWS, GitHub, Google, Google DeepMind, Microsoft, and OpenAI culminated in a $12.5 million funding initiative via the Linux Foundation to bolster open source AI security projects. This unprecedented alliance, rare among fierce competitors, acknowledged the pressing challenge posed by the flood of AI-generated vulnerability reports overwhelming maintainers. While this financial backing significantly expanded capacity—AWS alone contributed an extra $2.5 million to Alpha-Omega—it fell short of resolving the critical signal-to-noise dilemma, where maintainers struggled to discern valuable findings amid a sea of low-quality submissions.

The initial surge of AI-generated bug reports inflicted what Daniel Stenberg termed a 'denial of attention attack,' with projects like Curl ultimately terminating their paid bug bounty programs by early 2026 due to the distracting volume of low-quality submissions. However, by April 2026, a notable turnaround occurred as AI-assisted reports matured in quality, enabling maintainers to rely on them effectively. This evolution from 'completely useless to very useful' reflects a broader industry adaptation, underscoring that despite the noise, bug bounty programs remain indispensable for filtering truly exploitable vulnerabilities amid AI's prolific but often unimportant findings.

To manage the overwhelming influx of AI-generated reports, security teams have adopted strategic, phased approaches to bug bounty programs, starting with private, invite-only initiatives before gradual expansion. Shopify exemplifies this adaptive mindset by integrating AI-driven triage, validation, and reproduction agents into its open bug bounty program, enabling scalable processing and continuous learning from submissions. Viewing AI-enabled security researchers as the 'next wave of capability,' Shopify balances AI acceleration with traditional harness-building, resisting trends toward tiered payouts or program cancellations and instead innovating to maintain program openness and effectiveness.

Amid a fragmented social media landscape complicating bug reporting, the revival of Bug Track—a mailing list originally designed for centralized bug tracking—signals a community-driven effort to restore order and improve triage workflows. Human reviewers, often augmented by AI assistance, play a pivotal role in validating AI-generated bug reports, which remain only partially patchable by AI itself, with estimates indicating 53% cannot be accurately fixed by automated means. This resurgence also spotlights ongoing skepticism about relying solely on AI to audit AI-produced code, encapsulated in concerns over 'who watches the watchers,' emphasizing that human oversight remains indispensable in the evolving security validation ecosystem.

Sources
RockCyber MusingsThe PrimeTimeSecurity Weekly - A CRA ResourceEnterprise Security Weekly (Audio)Paul's Security Weekly (Video)

AI Arms Race Hits Crypto

AI-driven audits are rapidly exposing—and culling—vulnerabilities in blockchain ecosystems, but operational limits and human oversight remain vital as bots scour code at unprecedented scale.

By early 2026, AI-driven tools like EVM Bench began reshaping smart contract security by automating vulnerability detection and exploitation benchmarking, particularly targeting the long tail of smaller, low-value contracts on widely used platforms such as Ethereum's EVM. While well-established DeFi contracts retained relative safety due to extensive market testing, experts warned that AI's low inference cost would rapidly destabilize less mature protocols, as automated bots could economically exploit contracts with minimal locked value, effectively pruning insecure projects from the ecosystem.

The Bitcoin Red Team’s large-scale AI-assisted audits exemplify a new paradigm in ecosystem security, leveraging models like Moonshot AI’s Kimi K3 and OpenAI’s Claude to scan hundreds of Bitcoin-related repositories within hours, uncovering thousands of potential vulnerabilities—including 85 critical flaws in just 30 hours. Despite the impressive scale and speed, the team faced significant operational challenges, notably AI access restrictions imposed by providers such as OpenAI and Anthropic, which forced a pivot to Chinese open-weight models like Kimi K3 and GLM that offer fewer constraints but raise concerns about safety guardrails and control.

The Ethereum Foundation’s coordinated deployment of specialized AI agents across its protocol codebase marked a milestone in proactive vulnerability discovery, with agents assigned roles from reconnaissance to validation to systematically hunt bugs before adversaries can exploit them. This approach yielded real-world results, such as the identification and disclosure of CVE-2026-34219 in libp2p’s gossipsub layer, but also highlighted the indispensable role of human researchers who must sift through high volumes of false positives and rigorously reproduce findings to confirm true vulnerabilities, underscoring that AI augments rather than replaces expert security analysis.

The Bitcoin Red Team’s volunteer-driven, AI-assisted audit campaigns have not only demonstrated the transformative potential of AI in rapidly scaling security reviews across critical infrastructure but also exposed systemic tensions—such as reliance on foreign AI models due to restrictive US AI policies and the bottlenecks in triage and disclosure workflows. As Rob Hamilton and his team emphasize, while AI accelerates vulnerability detection, the ultimate security gains depend heavily on developer responsiveness and the establishment of continuous AI audit pipelines, especially for complex components like the Lightning Network, signaling a future where each project must integrate AI-assisted audits to remain resilient.

Sources
BanklessCryptoNews.netCoin Stories with Natalie BrunellValue The MarketsCRdecrypt

Humans Still Hold the Line

Despite AI's speed and scale, the flood of false positives and unreliable patches means skilled security professionals remain irreplaceable for triage, validation, and safe remediation.

By early 2026, it became clear that AI-driven security tools, while capable of rapidly discovering vulnerabilities—including subtle issues like missing authorization checks—lack true contextual understanding and reasoning, resulting in high volumes of findings with many false positives. As noted in analyses and case studies from entities like the Ethereum Foundation and the Bitcoin Red Team, AI-generated vulnerability reports often require extensive human triage and validation; for example, Ethereum researchers described that 'a candidate isn't a finding until there's a self-contained artifact that reproduces the failure against the real code.' This persistent need for human oversight underscores that AI tools serve as powerful assistants rather than replacements for skilled security professionals who interpret, prioritize, and communicate findings effectively to stakeholders.

Despite advances in reducing false positives—LLM-driven static analysis tools have achieved reductions in the 91–98% range—AI-generated patches remain highly unreliable, with studies from 1Password and Veracode revealing failure rates as high as 74%, and nearly half of AI patches introducing new vulnerabilities. This alarming statistic, highlighted by 1Password’s research and echoed by experts like Tim Jarret and Keith Hoodlet, reveals a dangerous false sense of security where incomplete or incorrect AI patches cause vulnerabilities to remain exploitable yet undetected. Consequently, human expertise remains essential to review, verify, and approve any AI-generated code changes before deployment, reinforcing the industry consensus that AI complements but does not replace human judgment in remediation workflows.

The rapid acceleration in vulnerability discovery powered by AI models such as Claude Mythos and Kimi K3 has outpaced traditional validation and patching workflows, creating a bottleneck where human capacity for triage, verification, and remediation coordination becomes the limiting factor. Industry leaders like Jay Chaudhry of Zscaler and Tom Gillis of Cisco emphasize that while AI can now comprehend complex legacy codebases and chain multiple vulnerabilities into high-impact attack paths, the operational challenge lies in managing this flood of findings with human-managed interim controls like Cisco’s Live Protect. This shift has led to a broad consensus that AI tools augment but do not replace skilled security professionals, who remain indispensable for prioritizing vulnerabilities, ensuring consistency, and maintaining effective security operations amid evolving AI capabilities.

Confidence in fully autonomous AI penetration testing has sharply declined, dropping from 29% in 2025 to just 9% in 2026, as organizations confront AI’s blind spots, inconsistent outputs, and the persistent need for human-in-the-loop approaches. Reports from Cobalt and experts like Derek Rush of Bishop Fox highlight that human judgment is critical to decide which leads merit pursuit and to reconstruct validated attack chains, a nuance AI alone cannot replicate. Moreover, operational challenges such as AI platform access restrictions—illustrated by the Bitcoin Red Team’s forced migration from OpenAI to Chinese open-source models—further emphasize the indispensable role of human expertise in adapting AI tools to real-world security environments and regulatory landscapes.

Sources
Paul's Security Weekly (Video)Application Security Weekly (Video)decryptCryptoNews.netBCCR

Toward Autonomous Security

AI-powered automation is transforming vulnerability management into a near-continuous, machine-speed process—forcing defenders to rethink patching and embrace real-time mitigation as the new normal.

By 2026, the security landscape is undergoing a profound transformation driven by AI-powered automation that promises to reshape vulnerability management into a near-autonomous process. Experts like Mandy foresee vulnerability management becoming almost unrecognizable within a decade, as AI agents not only identify vulnerabilities but also make judgment calls and initiate remediation actions, including patch creation and deployment. This evolution is critical as the volume and complexity of vulnerabilities surge due to the proliferation of interfaces, bots, and AI-generated code, with companies like Microsoft and 360 Digital Security reporting record-breaking numbers of discovered flaws, underscoring the need for continuous, machine-speed vulnerability management integrated directly into development pipelines.

Continuous AI-driven penetration testing is emerging as a game-changer in compressing the vulnerability discovery cycle from weeks to mere hours, enabling security teams to keep pace with rapid software development without sacrificing quality. Platforms like Expo, founded by the Microsoft Copilot team, exemplify this shift by autonomously simulating real-world attacks to discover and validate vulnerabilities at scale, providing clear attack paths and reducing false positives. This relentless AI-powered scanning approach not only offers defenders an asymmetric advantage over external researchers but also fuels cautious optimism about approaching 'unhackable' software, even as concerns persist about attackers gaining access to similar AI capabilities.

The acceleration of AI-driven vulnerability discovery has outpaced traditional remediation workflows, creating a widening gap that demands a fundamental rethinking of security operations. Industry leaders like Cisco and Zscaler highlight the urgency of adopting rapid mitigation technologies such as Cisco's Live Protect, which leverages eBPF for kernel-level shielding without disrupting production, alongside differentiated patching strategies tailored to system risk postures. As Jay Chaudhry warns, relying solely on patching is futile in an era where AI defines exploits within hours, necessitating a balanced security stack that integrates compensating controls like RASP and WAF while emphasizing detection and response over a patch-centric mindset.

The AI revolution is not merely adding a new layer to the security stack but fundamentally redrawing it, increasing complexity and risk as legacy systems coexist with emergent AI infrastructure layers. Mushtaba Hamid articulates this paradigm shift, noting the rise of open weight models and middleware such as agentic governance and policy controls that reconstruct how security applications are built and managed. This compounded complexity poses significant challenges for CISOs and security professionals, who must navigate an expanding and evolving attack surface while balancing the rapid automation of security workflows with robust governance to mitigate unpredictable outcomes inherent in AI-driven decision-making.

Sources
Paul's Security Weekly (Video)Eye on AISFCISO Talk by James AzarStartup to LastSecurity Now

Get the stories behind the trends

Deep-dive reporting and the weekly brief, in your inbox.