October 2026 Summaries
11 posts from Snyk
Filter
Month:
Year:
Post Summaries
Back to Blog
Snyk UNFILTERED, an invitation-only customer summit in San Francisco held before the launch of Evo Agentic AppSec and the AI Security Summit, convened security leaders for private, attribution-free discussions rather than presentations or product demonstrations. Participants focused on unresolved AI security challenges, including accountability when AI-generated code causes harm, the transition of AI from experimentation to a scrutinized business expense, and the difficulty of remediating vulnerabilities after they are identified. They also discussed how broad access to AI distributes security risk and responsibility across organizations, as well as the need for governance and non-bypassable controls that can support autonomous agents while recognizing the limits of human review. Snyk framed the discussions as part of a wider industry shift toward helping organizations innovate and operate securely at AI-driven speed, with customer feedback intended to inform its future products and community events.
Oct 28, 2026
705 words in the original blog post.
PHP2Shell refers to two critical unauthenticated remote-code-execution vulnerabilities in the PHP engine that were reportedly exploited before their public disclosure and quickly followed by widespread proof-of-concept availability, illustrating how infrastructure-level flaws can affect applications even when their code and dependencies are well managed. The discussion emphasizes that organizations need both internal software inventories and outside-in visibility into internet-facing systems, since many PHP deployments and version details may not be visible through package manifests, container scans, or exposed server banners. It argues that rapid response depends on having an up-to-date attack-surface inventory before an advisory appears, deploying targeted exploit checks within hours, applying mitigations or patches, and retesting to verify fixes. The post also notes a subsequent partial bypass and contends that AI-assisted vulnerability research and exploit development are shortening the interval between discovery, patch release, and active exploitation. It presents continuous external testing, heuristic scanning, and application-context reasoning as complementary capabilities for identifying both known vulnerabilities and more complex flaws, while promoting Snyk’s API and web security products as tools for this approach.
Oct 28, 2026
1,702 words in the original blog post.
Snyk has opened its Evo Remediation Agent preview, enabling AppSec teams to run recurring campaigns that automatically address open-source dependency vulnerabilities across selected repositories. Teams define campaign scope, severity targets, schedules, and optional non-breaking-change restrictions, while the agent plans upgrades, resolves dependency conflicts, validates changes in a hosted sandbox through builds, tests, and rescans, and submits documented pull requests through existing SCM workflows. Developers retain authority to review and merge proposed fixes, while AppSec teams gain visibility into campaign outcomes, unresolved issues, and remediation progress. Snyk reports that the engine achieved a 90.2% fully verified fix rate across 307 dependency-upgrade tasks in its VulnBump benchmark, compared with 77.5% for Claude Code using Snyk findings and 59.6% for Claude Code alone. The Open Preview is available for Snyk Open Source vulnerabilities in Evo Agentic AppSec, requiring workspaces, SCM integration, tenant preview access, and compliance with daily fix limits and public package registry use.
Oct 22, 2026
853 words in the original blog post.
PHP released versions 8.5.12, 8.4.27, 8.3.36, and 8.2.35 on October 21, 2026, to address two critical PHP runtime vulnerabilities, CVE-2026-109618 and CVE-2026-109619, each rated CVSS 9.2 and capable of enabling unauthenticated remote code execution in web-serving deployments. The flaws occur during the PHP request lifecycle before application code executes: one involves corrupted state after an out-of-memory condition, while the other involves stale request state reused by long-lived workers. PHP-FPM, Apache mod_php, CGI, FrankenPHP, LiteSpeed, and similar environments are affected, with mod_php described as especially exposed, while unsupported PHP 8.1 and earlier require vendor backports, source backports, or upgrades. The recommended response is immediate patching and verification of the PHP runtime actually deployed on hosts and container images, since Composer manifests and conventional dependency scans do not reliably identify interpreter versions. Temporary configuration changes may reduce exposure in limited circumstances but are not considered complete mitigations. Snyk explains that its Unmanaged, Container, API, and Web scanning capabilities can help locate affected installations or test exposed applications, whereas its Open Source scans generally cannot detect PHP runtime vulnerabilities from Composer dependency files alone.
Oct 22, 2026
1,571 words in the original blog post.
At Snyk’s AI Security Panel ahead of the 2026 AI Security Summit in San Francisco, representatives from Snyk, OpenAI, Google Threat Intelligence Group, and boldstart ventures argued that AI is lowering the cost and increasing the speed and scale of cyberattacks, exposing risks organizations had previously tolerated. Participants highlighted AI agents as a growing attack surface because they can access tools, credentials, code, and data autonomously, making identity controls, limited permissions, audit trails, and time-bound access increasingly important. They agreed that vulnerability discovery is accelerating faster than remediation, with organizational processes and human approval bottlenecks preventing many known fixes from reaching production. The panel advocated machine-speed defense, including automated triage, testing, remediation, and security operations, while reserving human involvement for exceptions, judgment, and intent-setting. Speakers differed on how nondeterministic AI systems should be supervised and how regulation should assign responsibility for agent-caused harm, but broadly supported standards, visibility, risk-based controls, and clearer accountability. Recommended priorities included using advanced models to test internal systems, improving the ratio of vulnerabilities fixed to vulnerabilities found, auditing agent capabilities, enforcing strong authentication, and reassessing whether uptime concerns are delaying necessary security fixes.
Oct 21, 2026
3,137 words in the original blog post.
Snyk’s 2026 State of Agentic AI Adoption, Volume III, analyzes anonymized scans of 3,023 enterprise environments and model-specific adversarial testing across 29 attacker goals to argue that model risk scores should guide investigation rather than serve as standalone evidence of exploitation. Among 1,693 environments with detected model identifiers, 74.5% referenced a model with at least one high-risk score and 49.2% referenced a critical-risk model, particularly for personal-data extraction and command execution, although only 37.5% of detected model occurrences could be matched to a score. The report emphasizes that an AI model’s practical risk depends on its deployment context, as agents and MCP servers can connect models to sensitive data, tools, and business actions; organizations with both agent indicators and MCP server-building code showed substantially more high-scoring model references. It recommends verifying which models, agents, servers, tools, and permissions are active, selecting alternative models when scores differ meaningfully by attack type, and relying on application-level safeguards such as least privilege, output validation, approvals, monitoring, and code scanning when weaknesses are common across models. The research also finds that lower-capability models tended to perform worse on several security measures, premium pricing did not consistently correlate with stronger security, open-weight models had higher average personal-data leakage scores in a small sample, and missing training-data lineage remains common, especially for proprietary models.
Oct 21, 2026
1,654 words in the original blog post.
AI-assisted development is increasing software security pressure by generating code faster than teams can review it, introducing vulnerabilities in areas such as business logic and access control, and enabling malicious dependencies that may compromise systems before conventional pull-request or CI scans run. Snyk argues that effective application security now requires prevention within coding workflows, deterministic CI scanning, and recurring application-wide analysis, combined with policies for blocking, remediation, human review, and independent fix validation. Its proposed Evo Agentic AppSec platform uses specialized security agents for threat assessment, code and secret scanning, dependency protection, remediation, and adversarial testing, while Snyk’s existing analysis engines provide contextual prioritization and verification. The company says this approach can help teams reduce vulnerability backlogs, prevent unsafe package installations, improve remediation quality, and retain human control over security policies, priorities, and approvals.
Oct 15, 2026
1,067 words in the original blog post.
Snyk’s VulnBump benchmark evaluates whether AI agents can remediate real vulnerable open-source dependencies while ensuring projects still install, build, and pass their existing tests. Across 307 upgrades in 43 repositories and six package ecosystems, Snyk found that its Evo Remediation Agent passed all verification gates in 90.2% of cases, compared with 77.5% for Claude Code supplied with Snyk’s recommended upgrade and scan, 59.6% for unaided Claude Code, and 46.9% for a simple package-manager version bump. Although agents generally preserved build and test success, the more difficult challenge was actually removing vulnerabilities, particularly in transitive dependencies and major version migrations; unaided Claude Code frequently reported fixes that rescans showed were incomplete. The results indicate that security intelligence, target-version guidance, independent rescanning, and iterative build-and-test repair loops substantially improve remediation outcomes, while major framework upgrades remain difficult for all evaluated systems. VulnBump is open sourced with its evaluation cases and emphasizes that agent-generated claims should be validated through independent scans, builds, and tests rather than accepted as proof of a successful fix.
Oct 15, 2026
2,176 words in the original blog post.
Snyk CTO Manoj Nair and Anthropic’s Alon Krifcher discussed how AI-enabled, autonomous attacks may reduce defenders’ response time, citing Snyk observations that new enterprise security findings are growing more than twice quarter over quarter and CrowdStrike data reporting a 27-second fastest breakout time. Nair argued that organizations should operate discovery, remediation, validation, and prevention simultaneously, prioritizing high-impact applications and testing them against realistic chained-vulnerability attack paths rather than simply retriaging backlogs. He highlighted Snyk products including Evo Continuous Offensive Security, remediation agents using Claude, and Snyk Studio, which provides security context to coding agents during development, while noting that AI-model safety guardrails do not address risks from credentials, configurations, packages, and other agentic components. Nair recommended tracking the ratio of new findings to resolved findings, along with the share of automated proposed fixes that merge without human rewriting, as measures of whether security programs are reducing risk rather than accumulating unresolved issues.
Oct 07, 2026
951 words in the original blog post.
A comparison on Snyk’s deliberately vulnerable TaintedPort application evaluated Evo Continuous Offensive Security (COS), which tested a live URL with optional source access, against Claude Security running Mythos, which analyzed source code only. Using a fixed benchmark of 57 known vulnerabilities and 15 exploit chains, Evo COS reported 50 vulnerabilities, confirmed 10 exploit chains, achieved 75.7% severity-weighted detection, and had a 91.7% F1 score, while Claude Security found 37 vulnerabilities, achieved 49.6% weighted detection, and had a 75.5% F1 score; Claude found one more critical-severity vulnerability. The comparison highlights an SSRF flaw and hardcoded JWT secret that all tools identified individually, while Evo COS reportedly demonstrated their use together to retrieve the secret, forge an administrator token, and access administrative functionality. The authors argue that dynamic testing can validate runtime behavior and multi-step exploitability that source analysis may only infer, whereas static analysis can reveal source-level logic and cryptographic issues that may not surface during live testing. They characterize the approaches as complementary, emphasizing that Evo COS combines multiple models, prior platform context, specialized agents, and independent validation, while noting the results came from one run per tool on an application built and maintained by Snyk.
Oct 07, 2026
1,143 words in the original blog post.
Broken access control, including broken object-level authorization (BOLA) and insecure direct object references (IDOR), remains a leading application security risk because authenticated users may access resources belonging to other users or tenants when ownership checks are omitted. The text argues that AI coding agents can readily generate such flaws because authorization requirements often depend on product-specific rules—such as organizational ownership, user roles, and intended tenant boundaries—that are not evident in a prompt or individual source file. While static application security testing can detect structurally recognizable issues within the broader access-control category, it generally cannot determine whether a particular object-level ownership check is required or correctly implemented. Effective detection therefore requires an application-wide model incorporating architecture, data flows, schemas, classifications, trust boundaries, and runtime behavior, combined with independent validation and testing. Recommended practices include inventorying endpoints that accept object identifiers, documenting resource ownership rules, explicitly reviewing authorization in agent-authored changes, adding cross-tenant functional tests that return indistinguishable not-found responses for unauthorized resources, and using contextual analysis and runtime testing alongside deterministic scans.
Oct 01, 2026
2,086 words in the original blog post.