IG5 (dsh-infinite-gen-5) bolts static analysis, native debugging, traffic capture, and agent collaboration onto DeepSeek Harness for authorized reverse engineering and defensive audits.
| Tool | Minglink/dsh-infinite-gen-5 — IG5, an AI-driven reverse-engineering plugin and visual workbench for DeepSeek Harness (DSH), v1.1.0 |
| Category | Reverse engineering workbench / LLM-agent analysis suite (JavaScript) |
| Primary Use | Authorized sample triage and deep analysis of software you own or are contracted to audit: decompilation, CFG inspection, x64dbg debugging, Unicorn emulation, HTTP/HTTPS capture, and decryption/protocol research |
| Safe Use | The README restricts use to self-owned or legally authorized targets for reverse engineering, security audits, and teaching/research; it explicitly prohibits use on systems without written authorization, DRM circumvention, and weaponization |
| Telemetry Note | Runs as a local plugin inside DSH with hash-verified installs; its proxy capture requires explicit TLS decryption setup and does not auto-modify system proxies or certificate trust, so defenders can observe it via local process and install-time SHA-256 manifest checks |
dsh-infinite-gen-5, branded IG5 at version 1.1.0, is not a standalone disassembler but a plugin and visual workbench layered on top of DeepSeek Harness (DSH). Its stated ambition is to merge static analysis, native debugging, traffic capture, data analysis, and multi-agent collaboration into a single workflow, which is an interesting architectural bet: rather than bolting an LLM chat window onto an existing RE tool, it treats the agent harness as the host and the analysis engines as pluggable backends. The repo carries over 3,400 stars, is written in JavaScript, and is explicitly non-commercial under CC BY-NC-SA 4.0.
The README is unusually disciplined about scope. IG5 exposes a default Core 8 toolset — ig5_doctor, ig5_open, ig5_status, ig5_funcs, ig5_strings, ig5_decompile, ig5_close, and ig5_profile — for daily analysis, with a Full 40 set available on demand. Crucially, those 40 entries break down into 25 read-only analysis tools, 13 modification-or-execution entries gated behind host approval, and 2 lifecycle/configuration entries. Any tool that renames, patches bytes, runs scripts, debugs, or emulates requires the DSH host to return allowed-once before it proceeds, and a denial, service unavailability, or ambiguous result means the action simply does not run.
Static analysis leans on Ghidra as the primary backend, and the full Windows package bundles Ghidra, a JDK, Python, x64dbg, bridge components, and upstream sources with license and hash manifests — no commercial IDA installation required. Capabilities listed include functions, strings, cross-references (ig5_xrefs), decompilation, CFG viewing (ig5_cfg), type and struct inspection, vtables (ig5_vtables), switch tables, microcode (ig5_microcode), and binary diffing (ig5_bindiff). The README is careful to note that support depends on the concrete backend, and that some analyses and unresolved results keep their pending markers rather than being silently resolved.
The visual workbench side renders SVG CFGs with pan and zoom, double-click drill-down to assembly, focused highlighting of local variables, struct drafts, and an audit timeline. The authors repeatedly qualify what these views mean: focus-line filtering is explicitly not full taint analysis, and draft edits still have to be sent through the approval pipeline. This habit of annotating limitations next to every feature is a hallmark of the whole document and, frankly, a good sign for a tool aimed at evidence-driven analysis.
Dynamic work is split between native debugging via x64dbg (breakpoints, registers, memory, restricted traces) and single-function emulation through Unicorn. The emulation lane is honest about not providing a full OS environment, imports, or TLS, which matches how professionals actually use Unicorn — for isolated function-level reasoning, not whole-program execution. There is also optional ScyllaHide integration documented separately, aimed at anti-debugging-heavy authorized targets, plus VMCI guest debugging and offline trace recovery for isolated lab setups.
Version 1.1.0, dated 2026-10-20, added independent HTTP/HTTPS capture via ig5_capture and ig5_capture_control, providing bounded evidence streams, traffic history, evidence export, request replay, and downloaded-payload extraction without depending on Burp. The boundaries are again explicit: only approved origins, no automatic modification of system proxy or certificate trust settings, and the README itself says it is not a full Burp replacement. Decryption and protocol work covers encoding and compression checks, XOR/AES, bounded candidate search, PCAP/PCAPNG parsing, framing, and observed state models, with no guarantees of unknown-key recovery.
The deobfuscation research layer is where IG5 gets philosophically interesting. It uses real p-code/SSA from Ghidra, single-function read-only deobfuscation overlays, path projection, state composition, and offline trajectory recovery. The research index documents VM-recovery components — observation layers, fetch-flow analysis, transition chains, opaque-effect projection — while stating plainly that partial models or mathematical proofs of specific joins do not equal generic unpacking, full VM recovery, or business-logic restoration. Each claim is graded: load success, IR generation, local equivalence, structural review, and real business recovery are accepted separately.
Agent collaboration is a headline 1.1.0 feature. Subagent tasks are dispatched natively by DSH, and IG5 implements a handoff → claim → evidence → submit → completion → review pipeline where formal review must bind to the same handoff ID and actual lifecycle; old reviews lacking a handoff_id count only as integrity pre-checks. A shared IG5 call budget across parent, child, and descendant sessions must be enabled explicitly, for example via ig5_profile collaboration={action:"start",max_calls:20}, and the budget cannot be inflated by repeated starts or reloads. The plugin verifies identity, evidence, and lifecycle — it pointedly does not certify that the model understood anything.
Installation is Windows x64 first: download the IG5-1.1.0-Windows-x64-full-20261020.zip plus its .sha256, verify the hash, then run plugin\install.ps1 from the extracted root, followed by /ig5 engines and ig5_doctor to confirm real availability. The installer verifies all assets, backs up existing plugins, and refuses to install on missing files or hash mismatch — and the README warns that a file merely existing does not mean the engine actually started. A source-tree install.ps1 exists for maintainers, with -Offline and -DistributionArchive/-DistributionRoot flags for reusing pinned donor assets, though the git-clone tree is not a complete offline package.
Platform coverage outside Windows is limited: an experimental darwin-arm64 Ghidra static lane with no x64dbg or working macOS Unicorn runtime, ARM64 target analysis but no finished mobile hosts, and a declared DSH range of @deepseek-harness >=0.2.0-rc.1 <0.3.0-0 with only 0.2.1-alpha.1 recorded as actually verified. Eight workflow skills — /ig5-triage, /ig5-deep-dive, /ig5-patch-and-sign, /ig5-diff, /ig5-debug-live, /ig5-crypto, /ig5-protocol, and /ig5-reverse-agent — package the tooling into repeatable analyst procedures, from first-look triage to persistent investigations with evidence preservation and per-item acceptance.
From a defensive-professional standpoint, IG5 is best understood as a governed evidence-collection layer for the agent-assisted RE era: every mutating action passes through host approval, every conclusion is tiered by verification level, and the license plus disclaimers confine use to owned or authorized targets, lab corpora, and education. For teams evaluating agent-driven analysis, the shared-budget accounting, stale-evidence marking on old attachments, and strict handoff binding are design patterns worth studying independent of the tool itself.
Minglink/dsh-infinite-gen-5.Educational analysis for authorized security professionals. Use only in controlled, authorized environments.
Related coverage
0 comentários:
Post a Comment
Note: Only a member of this blog may post a comment.