Opening Note
This week's developments converge on a central tension in security operations: AI agents are proliferating faster than the infrastructure and governance to manage them. Security practitioners are building custom agents because commercial products leave workflow gaps, yet Gartner predicts only 15% of SOC AI pilots will deliver measurable gains without structured evaluation. OWASP's OASIS initiative applies the same build-it-yourself impulse to open-source patch generation, keeping humans in the review loop. Meanwhile, attackers are exploiting trusted Microsoft infrastructure to assemble phishing pages inside browser memory, bypassing URL-based detection entirely. Meta's four-layer agent architecture offers one answer to the governance problem by separating knowledge from reasoning and regression-testing expert corrections. And Harness rebuilt its Git repository for nonstop agent traffic, acknowledging that coding agents are drowning review teams in pull requests. The Florida Politics ransomware column is excluded; its article body was absent from the supplied evidence.
Top Stories
Security Teams Build Their Own Agentic AI as Vendor Products Miss Gaps
Security professionals encounter AI agents everywhere. Nearly every vendor markets a product that uses them, and the term has become generic. The article notes that practitioners have started building, configuring, and operating their own agentic AI because custom-built agents fill the gaps that off‑the‑shelf tools cannot.
The author lists three reasons for the shift: agents work, custom‑built ones work better, and open‑source agents can strengthen collective defense. Open‑source, vendor‑neutral agents are presented as a way to share improvements across the community, but the piece warns that without proper governance they can cause more harm than good. The argument is presented without empirical benchmarks or case studies, so the claim remains unverified. Teams should therefore establish clear governance frameworks before deploying self‑built agents to limit unintended consequences.
Latent.space researchers log 20 billion tokens with GPT‑6 Astra
Latent.space reports that its engineers consumed over 20 billion tokens of GPT‑6 Astra while running exploratory workloads. The authors describe the effort as an operational deployment rather than a controlled benchmark, suggesting the observations reflect real‑world reliability and cost accumulation.
The report frames Astra as a hireable AI engineer, implying a labor‑substitution cost of under $6 per hour. No independent performance data are provided, and the article offers only a brief introductory statement. Because the evidence consists of a single passage from one team, the claim that Astra can serve as a cost‑effective engineering agent remains unverified. Practitioners should treat the cost figure as provisional and validate the model on specific tasks before relying on it for production work.
OWASP Introduces OASIS AI Pipeline to Generate Candidate Patches for Open‑Source Vulnerabilities
On August 26, 2026 OWASP announced OASIS, an initiative that applies artificial‑intelligence models to scan open‑source repositories and produce candidate patches for identified security weaknesses. The generated changes are then examined by application‑security professionals and validation agents to confirm that the vulnerability is genuine, the fix addresses the root cause, and no new regressions are introduced before the patch is offered to the upstream project. OWASP positions OASIS as a vendor‑neutral community effort rather than a commercial scanner, aiming to bridge the growing disparity between rapid vulnerability discovery and the limited capacity of maintainers to develop and test fixes.
The project’s design keeps final authority with the maintainers, who may accept, modify, or reject any AI‑generated suggestion. OWASP explicitly states that AI‑produced code will not be merged automatically; instead, it serves as a filtered starting point for human review. While the model promises to reduce the backlog of unfixed issues, the announcement provides no empirical data on patch acceptance rates, time‑to‑merge improvements, or the false‑positive frequency of AI‑generated patches. Adding an intermediary review layer could also increase coordination overhead for projects that already maintain robust security processes, and the quality of expert review may vary across a large participant pool. Practitioners should therefore treat OASIS output as supplemental triage material, integrating it into existing vulnerability management workflows while maintaining independent validation before any code is merged.
Further Reading
Phishing Pages Built Inside Browser Memory Using Blob URLs Evade URL Filters
The attack begins with a DocuSign‑themed email that carries a calendar invite. The invite points to a legitimate Microsoft OAuth endpoint, and a crafted redirect sends the user to Microsoft Teams. Teams loads a resource from an external domain (cdn.bloom.io). That resource creates a blob URL, a temporary browser‑generated address that references content stored locally in memory rather than on a web server. The malicious page registers a service worker and runs part of its logic inside a sandboxed iframe, allowing a remote backend to send live instructions through browser messaging. This architecture lets operators steer the phishing content per victim without relying on a static redirect.
Because the phishing page never resides on a remote server, conventional crawlers and link‑scanning tools that inspect only the first hop see only legitimate Microsoft URLs. Barracuda notes that the campaign removes many of the indicators security teams have traditionally used, pushing detection toward behavioral and identity‑based controls. The analysis is based on a single vendor source; no independent verification of campaign scale or active exploitation has been provided. Blob URLs and service workers are standard browser APIs, so detection rules that target them may generate high false‑positive rates unless they are tightly scoped to authentication contexts.
Meta unveils four‑layer agent architecture that stores knowledge in version‑controlled files and updates via regression‑tested expert edits
Meta’s design organizes more than 200 structured files—position, taxonomy, routing, and gateway files—into a strict taxonomy. The files feed composable reasoning recipes that load knowledge step‑by‑step. Routing indexes map inputs to the appropriate files without relying on embedding similarity, providing deterministic retrieval pathways. Expert corrections are captured at predefined checkpoints, root‑caused, compiled into minimal edits, and validated against replay and regression suites before being merged back into the knowledge base. This approach keeps institutional knowledge out of the model’s weights and enables permanent, regression‑tested improvements.
The architecture consists of four layers: a knowledge system that consolidates all institutional knowledge, a reasoning pipeline that separates what the agent knows from how it reasons, an evaluation framework that provides automated benchmarks, and a self‑improvement loop that updates the knowledge files without model retraining. Meta reports that the system reduced assessment time from days to minutes and experienced zero regressions across improvement cycles. However, performance claims are based solely on Meta’s internal assessment, and the design has been validated only in a single compliance domain. No independent benchmarking or scalability analysis has been published, and the maintenance overhead of a large file‑based knowledge base remains uncertain.
Practitioners considering this pattern should first audit a high‑risk workflow for separable knowledge and reasoning steps. Prototyping composable recipes with version‑controlled files and regression‑tested expert edits can reveal whether the maintenance effort aligns with operational constraints before scaling the approach to broader domains.
Gartner predicts 70% of large SOCs will pilot AI agents by 2028, but only 15% likely to see measurable gains without structured evaluation
Gartner analysts forecast that by 2028, 70% of large security operation centers will pilot AI agents to augment Tier 1 and Tier 2 functions, yet only 15% will achieve measurable improvements without a structured evaluation process. The prediction follows a shift in Gartner’s Hype Cycle, which now places AI SOC agents at the Peak of Inflated Expectations after a prior Innovation Trigger stage with single‑digit adoption. Independent survey data show that 40% of security teams already use AI daily and 56% are evaluating or piloting AI tools, indicating strong momentum in the market.
Gartner’s recommended evaluation framework starts from internal bottlenecks rather than vendor feature lists. It emphasizes four TDIR outcomes—mean time to detect, mean time to respond, false‑positive reduction, and mean time to contain—as the primary success criteria. The framework also requires real‑world benchmarks from environments comparable to the buyer’s and a clear distinction between proof‑of‑concept results and sustained production performance.
Evidence from recent surveys highlights operational challenges. In‑house AI tooling matched commercial products on initial investigation‑time gains, but 46% of those builds were abandoned because maintenance could not keep pace with evolving alert types and staff turnover. Additionally, 57% of respondents still require a human to review every AI decision before closing an alert, and no organization grants full, unsupervised autonomy to AI agents. These findings suggest that maintenance burden and limited automation scope are common failure points.
Practitioners should scope AI agent pilots against specific bottleneck workflows and track the four TDIR metrics from the outset. Establishing baseline measurements before deployment and demanding production‑level benchmarks from vendors can increase the likelihood of moving beyond the 15% success ceiling. A disciplined evaluation approach is essential to translate widespread AI adoption into tangible risk reduction.
Harness launches rebuilt Git repository and AI Code Review to ease PR overload from AI agents
Harness announced a rebuilt Code Repository and a new AI Code Review product a few days after a major GitHub outage. The offering targets the pull‑request bottleneck that engineering teams face as AI coding agents generate far more changes than human reviewers can handle. The AI reviewer works directly on pull requests that already live on GitHub, so teams do not need to migrate repositories to adopt the solution.
The core of the AI reviewer is a software‑delivery knowledge graph that maps a customer’s pipelines, deployments, incidents, and policies. This graph lets the reviewer fetch relevant context quickly and without consuming large numbers of LLM tokens. It surfaces substantive code changes first and separates routine scaffolding, such as dependency‑bump files, from changes that require human attention.
Harness reports that its customers saved more than 10,000 hours of manual review time each month after deployment. The company also notes that its repository runs on Kubernetes across multiple clouds, handles thousands of commits per second in testing, and is used by about 20 enterprise beta customers. These performance and savings figures are self‑reported; no independent benchmarks or published customer validations are available, and the beta customer outcomes remain unpublished.
Practitioners can reduce review bottlenecks by integrating the AI reviewer into existing GitHub workflows and using its change‑prioritization to focus human effort on substantive code modifications rather than routine scaffolding.
Closing Thoughts
Practitioners face a narrowing window to establish governance before ungoverned agents create more problems than they solve. The Gartner forecast should be read as a procurement warning: pilot volume will not equal outcomes. Teams building custom agents should adopt evaluation frameworks centered on MTTD, MTTR, and false-positive reduction before deployment, and should scope behavioral detection rules tightly to authentication contexts to avoid false positives from blob-URL phishing techniques. Watch for independent benchmarks of OWASP OASIS patch acceptance rates, third-party validation of Meta's architecture beyond compliance, and any replication of the browser-memory phishing technique across additional vendor research. The Harness and latent.space reports remain self-reported; treat their performance claims as provisional until customer validation surfaces.