
CCE Technical Intelligence Briefing — Volume 15
The AI governance thread is the most structured in any single volume to date. Heather Barnhart announces the SANS Digital Forensics + AI Investigations Framework and the Incident Response + AI Investigations Framework — the first formally published governance models that map AI use to investigation phases, risk levels, reversibility, and accountability requirements across the full forensic lifecycle. CyberTriage's architecture piece explains exactly how GenAI clients access investigation data through MCP, direct file access, and copy-paste — the clearest technical primer on the data flow question that practitioners have been asking since GenAI entered the forensic workflow. Rob Lee's third piece on the OpenAI/Hugging Face incident documents the investigation governance problem: OpenAI commissioned the independent review, chose the reviewers, set the scope to a seven-day window, and held redaction rights — raising the question of what an independent investigation actually means when the investigated party controls its parameters. Sumeshi's "Don't Make AI Your Forensic Analyst" provides the practitioner counterweight: a researcher who has built an experimental local-LLM forensic harness arguing from hands-on experience that small models misread instructions, lose long context, and repeat mistakes, and that the harness exists to compensate for exactly those failure modes.
The macOS storage thread is the most technically dense cluster in the volume. Six Eclectic Light Company pieces collectively document everything that can go wrong with APFS storage — cache accumulation and purging mechanics in Tahoe, how to take control of snapshots proactively, what happens when boot volumes genuinely run out of space, the specific cases where Finder's size on disk figure is meaningless, version versus snapshot versus backup as recovery mechanisms, and the full storage crisis management reference that Apple's own documentation omits. Logistician 1.4 adds improved charting to the log analysis workflow, and DeltaFree 1.1 improves the backup storage forecasting charts. XProtect 5358 adds six new YARA rules for Dubrobber and ShadyShoelace families.
The LEAPPs thread is exceptional. Three Brignoni pieces cover VLEAPP's revival to read complete vehicle head unit data beyond a single ECU, the Powerlog timestamp investigation that discovered a 1971 clock bug in how iOS reports power events, and MMKV — the Tencent key-value storage library used by WeChat and many other major apps — and what it stores forensically. These three pieces collectively advance the vehicle forensics and mobile artifact research agenda more than any single volume in the digest's history.
The wellbeing thread this volume is anchored by Paul Gullon-Scott's most forensically specific piece yet — a clinical dismantling of police recruitment language that frames resilience as the price of entry into digital forensic work. The evidence Gullon-Scott marshals is precise: cumulative exposure overwhelmed investigators who started out resilient, psychological distress worsened over 18 months in the longitudinal study, and established positive coping strategies did not significantly reduce secondary traumatic stress in the sample. The S21 VisionX victim identification piece and the uncomfortable question about evidence visibility are the CSAM investigation complement — the operational argument for why tools must surface what they cannot process rather than silently hiding it.
The hardware and tooling thread features Atola Boot Image — a free bootable environment that solves the soldered-drive acquisition problem by exposing internal drives as write-protected iSCSI sources rather than requiring physical removal. Elcomsoft Quick Triage 2.2 adds timeline, file system snapshot, and plugin engine capabilities. Matthew Plascencia's three-way tool comparison (Belkasoft X, Oxygen Forensic Detective, ADF Pro) and the Lucid Truth share link forensic artifact reference complete the methodological toolkit thread. ThinkDFIR's Amcache piece is the discovery of the volume — SQLite databases in the Windows AppCompat folder that appear to be replacing the Amcache registry hive, undocumented and with no obvious changelog entry.
Andrea Fortuna
Incident Response Is a Business Decision, Not an IT Ticket
Fortuna uses a COO asking during an active incident whether it was safe to keep the production line running — and not getting an answer fast enough — as the entry point for the argument that NIS2 has just made legally binding what has always been operationally true: incident response decisions are business decisions made under evidence uncertainty, not IT determinations made when the forensic picture is complete. The gap between what the evidence supports and what the business needs to decide is where most incident response actually happens. NIS2's 24-hour early warning and 72-hour reporting requirements do not wait for forensic certainty; they impose legal obligations on decisions that must be made with partial information. For forensic practitioners whose organizations operate under NIS2, this piece provides the framing for why their findings must be communicated in business decision terms from the first hour, not after the investigation is complete.
Andrea Fortuna
Weekly Wire #8: When Trust Chains Break
Fortuna's eighth Weekly Wire covers the week in which 153 million driver's licence images were traced back to rental counters and dispensary scanners by Brian Krebs, ClickFix operators discovered that Polygon's blockchain immutability makes it a better C2 layer than any domain they could register, and SonicWall's edge appliances — designed to guard the perimeter — became the perimeter's softest spot. The common thread across all three is the same structural failure: infrastructure that was trusted because it was trusted before, breaking in the direction of the adversary. The velocity at which the gap between 'this infrastructure is trustworthy' and 'this infrastructure is the breach' is shrinking is the operational observation that matters for incident responders.
Belkasoft
What's New in Belkasoft X v2.12
The full release notes for Belkasoft X v2.12, providing the complete feature list beyond the sneak peek covered in Volume-14. Key additions include offline evidence translation with BelkaGPT for chats, messages and artifact text; BelkaGPT-powered SQLite database classification with predefined and custom categories; custom SQLite artifacts for structuring and analyzing data from unsupported databases; the BelkaGPT Hub installer for Windows for easier distributed offline AI processing deployment; redesigned MTK workflow for Android acquisition; and a Belkasoft Offline Maps feature for location evidence analysis without internet connectivity. For labs that evaluated the sneak peek, this is the definitive feature reference for the 2.12 release.
Elcomsoft
Elcomsoft Quick Triage 2.2: Timeline, File System Snapshot, and a Plugin Engine
Elcomsoft Quick Triage 2.2 adds three capabilities that significantly expand its analytical utility. First, a Timeline view that merges events from every timestamped artifact into one chronology — browser history, Prefetch, USB registry keys, and all other time-stamped sources normalized to a single axis and sorted, eliminating the mental overhead of interleaving multiple tables. Second, a file system snapshot that copies the metadata tree (directory structure, file names, sizes, timestamps, hashes) without copying file contents — enabling rapid baseline documentation without the storage overhead of a full image. Third, a plugin architecture for artifact parsers that allows practitioners to extend the tool's coverage to new artifact types through external modules. MSA password attacks, OpenDocument parsing, and recursive archive parsing in full-text search are additional additions.
Cellebrite
Cellebrite Federal Solutions Achieves CMMC Level 2 Accreditation
Cellebrite Federal Solutions, Inc. (CFSI) achieves CMMC Level 2 accreditation — the advanced cybersecurity tier required for contractors handling Department of War Controlled Unclassified Information. The accreditation, assessed by a Certified Third-Party Assessment Organization, covers both technical safeguards and the policies, procedures, and demonstrable operational maturity required for the standard. For US federal agencies evaluating forensic platform vendors against DoW contracting requirements, CMMC Level 2 accreditation is now a documented baseline for CFSI.
Cellebrite
Digital Evidence Overload: Why Prosecutors Need Direct Access to Case Data
Cellebrite makes the case that prosecutors waiting for final forensic reports — rather than having direct access to the evidence repository throughout the case — is a structural bottleneck that weakens charging decisions and courtroom preparation. The practical argument is that static, summarized reports cannot support follow-up questions, contextual exploration, or real-time courtroom evidence recall the way direct access to a secure, audited evidence platform can. The piece documents the specific friction points that direct access removes: repeated examiner consultations for follow-up questions, delayed corroboration of investigative leads, and the inability to verify report accuracy against original evidence before trial.
Cellebrite
Cellebrite Guardian: IRAP-Assessed Cloud Platform Achieves PROTECTED Level for Australian Government Agencies
Cellebrite Guardian completes its IRAP assessment at the PROTECTED classification level — the independent evaluation framework required for cloud platforms used by Australian government agencies operating under the Information Security Manual and Protective Security Policy Framework. The assessment was conducted by CyberCX, an ASD-endorsed IRAP assessor, producing a Cloud Controls Matrix, a Cloud Security Assessment Report, and control recommendations. For Australian law enforcement, defence, and national security agencies that have been unable to adopt cloud-based digital evidence management due to PROTECTED classification requirements, Guardian's IRAP assessment removes the primary compliance barrier.
The Eclectic Light Company
DeltaFree Version 1.1: Greatly Improved Charting
DeltaFree 1.1 revises the point and line plot display substantially — rewriting the Swift Charts-based visualization layer to be easier to interact with and more analytically useful for the backup storage forecasting task. The data and file format are unchanged from version 1.0, so existing data files are fully compatible. For forensic labs that installed DeltaFree 1.0 for backup storage monitoring, this is a drop-in update that improves the display of the time-series data already being collected.
The Eclectic Light Company
Caches and Purging in macOS Tahoe
A systematic reference on how macOS caches accumulate and how the OS purges them when storage pressure builds — covering QuickLook thumbnail caches, Safari and Mail caches, app-managed caches, and the APFS purgeable space mechanism that allows the OS to reclaim space from Time Machine snapshots and cache data without explicit user action. The forensic significance is in the purging behavior: cache purging is asynchronous, triggered by storage pressure rather than by user action, and happens in the background during low-activity periods. For examiners interpreting free space calculations and cache artifact presence on examined systems, understanding the trigger conditions and timing of automatic purging is essential for accurate characterization of what was present at what time.
The Eclectic Light Company
Apple Has Released an Update to XProtect for All macOS (v5358)
XProtect version 5358 adds six new YARA rules for the Dubrobber and ShadyShoelace malware families — MACOS.DUBROBBER.SAINDY, MACOS.DUBROBBER.CHREIM, MACOS.SHADYSHOELACE.OSLOSTA, MACOS.SHADYSHOELACE.OSLODL, MACOS.SHADYSHOELACE.LODLSYDEAR, and MACOS.SHADYSHOELACE.LODLSYDEX8 — plus two new Osascript rules for MACOS.ANGRYORB.JXPL and MACOS.OSASCRIPT.DUEXKE. The SHADYSHOELACE family was first documented in XProtect 5357; this update substantially extends its coverage. For forensic examiners documenting XProtect version at examination time, 5358 is the current baseline as of early September 2026.
The Eclectic Light Company
Version, Snapshot or Backup: Which Is Best for Reverting Edits?
A structured comparison of macOS's three mechanisms for reverting files to earlier states — app-managed versions (Revert To in supported apps), APFS snapshots (Time Machine local), and Time Machine backups — covering how each works, what it preserves, how far back it reaches, and the specific failure modes of each. The forensic relevance is in the evidence availability assessment: a file's recoverable history depends on which of these mechanisms was active on the system at examination time, and understanding which mechanism preserved a given version is prerequisite to characterizing the provenance of recovered content accurately.
The Eclectic Light Company
Logistician 1.4: Improved Charts and Plots
Logistician 1.4 applies the same chart improvement work done in DeltaFree 1.1 to the log statistics visualization — upgrading the three point plot windows that display log daemon activity over time to be more interactive and analytically useful. For forensic practitioners who use Logistician to establish Mac logging activity baselines, this version improves the visual interface for the data that was already being analyzed without changing the underlying log statistics parsing.
The Eclectic Light Company
When the Finder's Size 'On Disk' Is Misleading
Using virtual machine bundle duplicates as the demonstration case, this piece documents the specific APFS clone file scenario where Finder's Get Info 'size on disk' figure is not just imprecise but actively meaningless — reporting the full allocated storage blocks for a clone that shares those blocks with its source, producing a size figure that would sum to more than the total container capacity if all clones were counted. The forensic significance is direct: disk usage calculations on APFS volumes that contain clone files — common in forensic lab environments that duplicate VM images — cannot be based on Finder's size-on-disk figures without risk of substantial overestimation.
The Eclectic Light Company
Last Week on My Mac: Storage Crisis Management
A comprehensive storage management reference that covers everything Apple's own documentation omits: snapshot management, cache purging mechanics, the difference between purgeable and actually recoverable space, and the specific sequence of steps for diagnosing and recovering from a genuine storage crisis. The ClickFix warning embedded in the piece is directly relevant to forensic practitioners: Google searches for Mac storage recovery advice are actively serving ClickFix attacks, making this a piece worth sharing with forensic lab users who might otherwise turn to search results for storage troubleshooting.
The Eclectic Light Company
What Happens When a Mac's Boot Volumes Run Out of Space?
Based on controlled observations in a macOS 26.6.2 VM with a 100 GB container, this piece documents the visible effects and log entries produced as APFS free space approaches zero — the sequence of system behaviors from initial performance degradation through the specific failure modes that become critical when the Data volume and VM volume have insufficient space for normal operation. For forensic examiners who investigate Macs that experienced storage crises, understanding the log signature of a storage-full condition provides a diagnostic anchor for reconstructing whether a system failure, data loss, or behavioral anomaly was storage-related.
The Eclectic Light Company
Take Control of Snapshots
A comprehensive reference on APFS snapshot management covering what snapshots are, how Time Machine creates and uses them, how much storage they consume and why, how to list and delete them, and how to use them for rollback. The forensic significance lies in the storage accounting section: a snapshot retains every storage block that was in use at the time it was taken, meaning that files deleted after a snapshot was created are not actually deleted from the device's storage until the snapshot is itself deleted. For examiners working Mac devices where storage consumption is anomalously high or where 'deleted' files should be unrecoverable, snapshot presence is a primary alternative explanation.
Forensic Focus / Belkasoft
Belkasoft X 2.12 Introduces Offline Translation for DFIR Investigations
The Forensic Focus announcement of Belkasoft X 2.12 highlights the offline translation capability as the headline feature — BelkaGPT can now translate text from chats, messages, and other artifact text without internet connectivity, directly addressing the air-gapped investigation environment that cloud translation services cannot reach. The piece provides a practitioner-level summary of the 2.12 feature set for readers who want a concise overview rather than the full Belkasoft release notes.
LEAPPs / Alexis Brignoni
The VLEAPP Revival: Reading the Whole Head Unit
Brignoni documents the revival and substantial expansion of VLEAPP — the vehicle artifact parser in the *LEAPP family — to read the complete data picture from modern automotive head units rather than a single ECU. The revival addresses a coverage gap that had persisted since VLEAPP's initial release: modern vehicles have multiple interconnected ECUs and infotainment systems, and parsing only one left the majority of the vehicle's digital evidence unread. The expanded VLEAPP now handles the multi-source artifact landscape of contemporary head unit data including navigation history, call records, paired device information, and application data across the full head unit architecture. For practitioners working vehicle forensics cases, this represents a significant expansion in the artifact coverage available through the open-source LEAPP toolchain.
LEAPPs / Alexis Brignoni
Powerlog Timestamps, Offsets, and a Clock That Thinks It's 1971
Brignoni documents the discovery that iOS Powerlog timestamp offsets are not always applied correctly by parsing tools, producing timestamps that appear to be from January 1971 rather than the actual event date. The investigation traces the error to the interaction between Powerlog's internal timestamp format and the offset field that adjusts for timezone and epoch differences — a combination that produces a plausible-looking but incorrect timestamp when the offset is not handled correctly. For examiners who use Powerlog data for iOS timeline reconstruction, this piece is the essential reference for verifying that the parsing tool being used handles the offset correctly rather than silently producing 50-year timestamp errors.
LEAPPs / Alexis Brignoni
You Down with MMKV?
Brignoni documents MMKV — Tencent's open-source key-value storage framework used by WeChat, and increasingly by a wide range of major iOS and Android applications — as a forensic artifact source that most practitioners have not yet systematically addressed. MMKV stores data in a memory-mapped file format that differs from SQLite and plist, requiring specific parsing tools for reliable extraction. The piece covers what MMKV stores, where to find it on both iOS and Android devices, the forensic significance of the data it typically holds, and the current state of LEAPP-based parsing support. For examiners who encounter WeChat or other MMKV-using applications in their casework, this is the foundational reference for the artifact class.
Magnet Forensics
Digital Fraud in a Connected World: How It Works and How to Prevent It
Drawing on FBI data showing digital fraud losses exceeding $16 billion in 2024, Steve Gemperle documents the social engineering mechanics that make digital fraud effective — specifically the urgency-and-isolation combination that prevents victims from verifying before acting — and the AI acceleration that is lowering the barrier to convincing fraud at scale through polished phishing, deepfake voices, and synthetic documents. The piece's most practically useful observation for forensic practitioners is the attribution challenge: AI-generated fraud artifacts (synthetic voice calls, deepfake video, AI-written phishing) require different evidence sources and different analytical approaches than traditional fraud investigation, and the fraud investigation community has not yet fully developed the methodological response to AI-generated evidence.
Matthew Plascencia
Belkasoft X vs Oxygen Forensic Detective vs ADF Pro: A Three-Way Comparison
Plascencia documents a head-to-head test of three commercial mobile forensic platforms against the same extraction — Belkasoft X, Oxygen Forensic Detective, and ADF Pro — structured around the specific scenario where a single piece of data is not easily parseable manually and the comparison reveals which platform handles it correctly. The piece is valuable precisely because it is structured around a failure case rather than a feature comparison: which tool correctly handles the edge case that matters to the investigation. For practitioners evaluating platform selection or understanding the gaps in their current toolkit, this provides the comparative methodology reference rather than vendor-supplied feature tables.
Rob Lee
Investigation Template Defined for AI Accidents: Hugging Face and OpenAI Updates and New Risks
Rob Lee's third piece on the OpenAI/Hugging Face incident identifies the most consequential aspect of what was published on August 26: not what the agents did, but how the investigation was structured and what that structure means for AI accident accountability going forward. OpenAI commissioned the independent review, chose the reviewers, set the scope to a seven-day window, held redaction rights over the output, and gave feedback on structure, emphasis, and tone. The reviewers were not paid by OpenAI but disclosed that the relationship created potential bias concerns. Lee argues that this investigation structure — commissioned by the investigated party, scoped by the investigated party, redacted by the investigated party — is not an independent investigation by any forensic standard, and that the precedent it sets for future AI accident investigations has direct implications for the credibility of the AI safety field's own accountability mechanisms.
Thesis Friday
Thesis Friday #27: Backward Reasoning from a Provable Endpoint
Principle 4 of Anchored Log Reconstruction reverses the direction of log analysis: instead of searching for what you expect to find (which nearly always produces a hit, making the finding worthless), start at a state the log can prove and work backwards toward a cause. A modern system writes so much that you can find support for almost any account you bring to it — forward searching confirms hypotheses without testing them. Backward reasoning from a provable endpoint (a confirmed login, a documented file creation, a verified network connection) establishes what must have been true before that endpoint occurred, which is a structurally different and more defensible analytical claim. For examiners building Unified Log-based reconstructions, this principle is the methodological antidote to confirmation bias in log analysis.
ThinkDFIR
Are We Going to Stop Calling It Amcache?
While exploring the Windows AppCompat folder, the author discovers that C:\windows\appcompat\programs now contains SQLite databases that appear to correspond to the sections previously found only in the Amcache.hve registry hive — with schemas that are structurally similar and include a LastModified FILETIME timestamp field that appears to replace the registry last write time used for forensic dating purposes. The change is undocumented: no version number has been identified, it does not appear in obvious update history or changelogs, and the exact Windows build in which it appeared is unknown. For forensic examiners whose Amcache workflow is built around the registry hive, this piece is a heads-up that the artifact may be migrating format — with immediate implications for tool compatibility and examination procedures.
Lucid Truth Technologies
Share Link Forensic Artifacts: What a Share Link Records, and What It Doesn't
Kenneth Hartman documents the forensic artifact landscape for share links — the URL-based sharing mechanisms used by Google Drive, OneDrive, Dropbox, Box, and similar platforms — covering what the generating platform records, what the receiving platform records, and what each establishes evidentially. The core distinction is that a share link records a permission grant event, not an access event: the link was created and shared, but whether the recipient actually followed it requires access log evidence from the platform rather than the link itself. For forensic practitioners and attorneys who have been treating share link presence as evidence of access rather than evidence of sharing, this piece documents the specific evidence required to establish each evidentiary proposition.
Sumeshi / GitHub
Don't Make AI Your Forensic Analyst
Sumeshi — a DFIR researcher who built forensia, an experimental local-LLM harness for Windows digital forensic investigations — argues from hands-on experience that AI should not be positioned as the forensic analyst in an investigation workflow. The piece documents the specific failure modes that motivated forensia's design: small models in the 4B–8B range misread instructions, lose long context across long artifact sets, and repeat mistakes. The harness exists to compensate for those failure modes by treating the LLM as one narrow component of an investigation loop rather than as an autonomous analyst. The forensic implication is structural: AI models that handle only one narrow step at a time, with a harness that checks results against evidence and controls what may be concluded, are more forensically defensible than AI systems that operate across the full investigation scope without those guardrails.
Heather Barnhart / Smarter Forensics
AI-Assisted. Human-Led. Why DFIR Needed Its Own AI Frameworks
Heather Barnhart announces the SANS Digital Forensics + AI Investigations Framework and the Incident Response + AI Investigations Framework — both free from SANS — as the first formally published governance models that map AI use to forensic investigation phases with explicit risk levels, reversibility assessments, and accountability requirements. The frameworks establish three human oversight tiers: human in the loop (approve before execution), human on the loop (monitor in real time, can intervene), and human over the loop (governance and policy boundaries). Non-negotiable lines are explicit regardless of AI capability improvement: AI never independently determines attribution, never makes legal or evidentiary conclusions, and every AI-assisted action is documented with transparent records. Barnhart identifies the biggest risk as not AI itself but uncritical trust in AI output — confidence without validation — and recommends starting with low-risk use cases, building governance, then expanding in that order.
Atola Technology / Forensic Focus
Atola Boot Image Enables Forensic Acquisition Without Removing Internal Drives
Atola Technology releases Atola Boot Image — a free bootable USB environment that solves the growing problem of forensic acquisition from laptops, workstations, and servers with soldered or otherwise difficult-to-remove storage. The environment write-protects the target computer's internal drives on boot, exposes them as iSCSI targets (with drive model and serial number in the IQN), and displays the IP address needed to connect from the forensic imager — allowing acquisition through Atola TaskForce 2, TaskForce, or Insight Forensic without opening the device. The approach differs from traditional bootable forensic environments like PALADIN and CAINE: rather than turning the target computer into the forensic workstation, it exposes protected drives to the examiner's existing imaging system, keeping acquisition, hashing, reporting, and other operations within the dedicated forensic imager.
Semantics21 / Forensic Focus
Semantics21 Asks the Uncomfortable Question Every Review Team Should Be Asking
Semantics21 frames the evidence visibility question that most forensic review teams do not explicitly ask: when your tool reports a media count and a processed case, is everything actually visible to the investigator, or have corrupt, damaged, and unsupported files been excluded and moved out of the working review view after initial processing? The piece argues that partial visibility creates false completion — a workflow that looks organized and appears to be working while potentially hiding investigative-relevant material inside the excluded set. A corrupt file is not automatically worthless: damaged images may carry EXIF metadata, broken videos may contain device information and naming patterns, and files that cannot be fully opened may still explain provenance. S21 VisionX's approach — keeping all files visible to the investigator regardless of processing status — is the practical response to the false-completion problem.
Semantics21 / Forensic Focus
How S21 VisionX Helps Investigators Look for the Victim Behind the Category Label
Semantics21 makes the operational distinction between categorisation and victim identification — a distinction that matters most in CSAM investigation where the category label answers what the material depicts but not where the child is, who they are, or what visual and contextual details in the material could move a victim identification workflow forward. Categorisation organises the material; it does not surface the school crest, the background location, the identifiable object, or the visual pattern that might represent the next investigative lead. S21 VisionX's victim identification workflow is designed to surface those details at scale for human assessment rather than treating categorisation as the end of the analytical process. The uncomfortable question the piece poses directly: does your chosen tool care as much about protecting children as you do?
Paul Gullon-Scott / Forensic Focus
The Skills and Resilience to Do One of the Hardest Jobs in Policing: Why This DFIR Recruitment Language Gets the Evidence Wrong
Gullon-Scott subjects a UK police digital forensic recruitment post — inviting candidates to join 'if you have the skills and resilience to do one of the hardest jobs in policing' — to the same evidence standard he applies to any forensic claim. The longitudinal study data is precise: psychological distress worsened over 18 months as cumulative exposure increased in investigators who initially appeared to be coping well; established positive coping strategies did not significantly reduce secondary traumatic stress in the sample; resilience at recruitment did not guarantee resilience years later. The recruitment language locates the ability to cope entirely inside the individual, letting the organisation off the hook for workload management, supervisory awareness, and structured support — the factors the evidence identifies as most protective. Gullon-Scott is not arguing against honesty about the difficulty of the role; he is arguing that framing resilience as the price of entry reinforces the stigma that already prevents struggling investigators from seeking help, and that recruitment language from a head of forensics shapes the cultural environment that the evidence identifies as the most important protective factor.
CyberTriage
How to Let AI Access Your DFIR and SOC Investigation Data
Dr. Brian Carrier provides the clearest published explanation of the GenAI client-server architecture as it applies to digital investigation data — the technical foundation that practitioners need to understand before making any decision about where their case data goes when they use AI tools. The key distinction is between the GenAI client (where the user interacts and where data integration happens) and the GenAI server (where the LLM processes requests): only a local client like Claude Desktop or Code can access local files and MCP servers; a web-based client has no pathway to local data beyond copy-paste. The three data access methods — copy-paste/upload, MCP integration, and direct file access — are documented with their specific read-only versus write-access implications for evidence integrity. The Autopsy and Cyber Triage MCP servers use MCP over STDIO, which requires no network ports and gives Claude Desktop access to investigation data through tool descriptions that guide the AI toward relevant data rather than requiring the analyst to upload raw artifacts.