029 — the threat model, package-wide; JSON nesting bounded before the decoder on both transports by HackTuah · Pull Request #40 · ScriptKittyOS/beam_mcp

029 red: JSON nesting is bounded before the decoder runs, on both transports, and docs/threat-model.md cites a test per refusal -- measured first: a 1 MiB body nested 524 288 deep is decoded in ful…

HackTuah

committed

1b196e3

029 green: BeamMCP.JSON -- the one place the wire's JSON is read on both transports, a byte walk in front of the decoder refusing a body nested past 64 levels by name (-32600; 400 with the connecti…

HackTuah

committed

bc4a429

029 mutants: ten over the nesting scan and the two transport clauses -- the walk ignored, off by one, objects uncounted, no count-down, strings not skipped, escapes not honoured, the wrong depth na…

HackTuah

committed

09696c6

029 round 1 red, lane a's findings: the bound pinned by bytes (64 admitted, 65 refused, literal -- a lane moved the constant to 8 and the suite stayed green) and by the page's exact phrase; a repea…

HackTuah

committed

a13bb7a

029 round 1 green, lane a's findings: a repeated key refused by name on both transports through Jason's ordered objects read once into maps (a key compared after unescaping, so "a" and "\\u0061" ar…

HackTuah

committed

34c79c1

029 round 1 mutants: eight over the repeated-key walk and the stdio answers -- last-wins instead of refusal, a list handed back as decoded, the wrong key named, the fault answered without the id, t…

HackTuah

committed

43599e3

029 round 1 red, lane b's findings: the wire table's REFUSED and BOUNDED rows each cite a test the reader can see (the repeated-key row's citations vanished behind escaped quotes in its prose colum…

HackTuah

committed

b409563

029 round 1 green, lane b's findings: the citation reader scans from the line's first backtick path, so a row's own quotes cannot swallow the citations (the repeated-key row's three had vanished), …

HackTuah

committed

5063763

029 round 1 mutants, lane b's: seven over the line cap's comparison, the legacy drain (nothing drained, one chunk, the wrong frame named) and the citation reader (the whole line scanned, a skipped …

HackTuah

committed

35aab76

029: Mjs20 re-anchored twice -- an unused drain and then an unused variable were compiler kills; drains zero bytes with the length kept in use, killed by the legacy-frame test

HackTuah

committed

1a8d8b3

029 round 2, lane a's four: the moduledoc's cost band is the page's (2.0-2.6x, not 1.0); the citation reader sees every ExUnit skip spelling -- a tag or describetag naming skip among the attribute …

HackTuah

committed

dfd478f

029 round 2 mutants: the reader reading one attribute line, the moduletag ignored -- zero survivors, one re-anchored past a compiler kill

HackTuah

committed

10a840e

029 round 2 red, lane b's findings: a header line of a legacy Content-Length block bounded like any line (today read unbounded -- 64 MiB accepted); the loop's heap about the line's bytes (today 92 …

HackTuah

committed

3888013

029 round 2 green, lane b's findings: every header line of a legacy Content-Length block is read under the line bound, and past it the block and its declared body are drained (64 MiB on one header …

HackTuah

committed

8d279bd

029 round 2: the heap pin samples the loop while it reads (what it retains after is nothing under either design, so that was no pin), and found a second forty -- the legacy Content-Length check dow…

HackTuah

committed

67bc6c6

029 round 3, lane a's low: the nesting row's heap figures are MiB as labelled -- 6,508 KB and 10,907 KB are 6.4 and 10.7 MiB, not the decimal 6.5 and 10.9 that stood under the label; the ratio 1.7x…

HackTuah

committed

41fc637

029 round 4, lane b: the nesting row's heap sentence compared a sampled peak with a resting heap and called the quotient 1.7x -- like for like, the bounded decode's sampled peak on the 60,000-key o…

HackTuah

committed

a54d6b1

029 round 5, lane a: the after-return heap clause dropped -- the two lanes read it two ways, because what a process holds after a decode returns is when it last collected, not a property of the dec…

HackTuah

committed

e665a26

Navigation Menu

029 — the threat model, package-wide; JSON nesting bounded before the decoder on both transports - #40

029 — the threat model, package-wide; JSON nesting bounded before the decoder on both transports#40
HackTuah merged 18 commits into
mainfrom
slice/029-threat-model

Commits on Sep 16, 2026

029 red: JSON nesting is bounded before the decoder runs, on both transports, and docs/threat-model.md cites a test per refusal -- measured first: a 1 MiB body nested 524 288 deep is decoded in ful…

029 green: BeamMCP.JSON -- the one place the wire's JSON is read on both transports, a byte walk in front of the decoder refusing a body nested past 64 levels by name (-32600; 400 with the connecti…

029 mutants: ten over the nesting scan and the two transport clauses -- the walk ignored, off by one, objects uncounted, no count-down, strings not skipped, escapes not honoured, the wrong depth na…

029 round 1 red, lane a's findings: the bound pinned by bytes (64 admitted, 65 refused, literal -- a lane moved the constant to 8 and the suite stayed green) and by the page's exact phrase; a repea…

029 round 1 green, lane a's findings: a repeated key refused by name on both transports through Jason's ordered objects read once into maps (a key compared after unescaping, so "a" and "\\u0061" ar…

029 round 1 mutants: eight over the repeated-key walk and the stdio answers -- last-wins instead of refusal, a list handed back as decoded, the wrong key named, the fault answered without the id, t…

029 round 1 red, lane b's findings: the wire table's REFUSED and BOUNDED rows each cite a test the reader can see (the repeated-key row's citations vanished behind escaped quotes in its prose colum…

029 round 1 green, lane b's findings: the citation reader scans from the line's first backtick path, so a row's own quotes cannot swallow the citations (the repeated-key row's three had vanished), …

029 round 1 mutants, lane b's: seven over the line cap's comparison, the legacy drain (nothing drained, one chunk, the wrong frame named) and the citation reader (the whole line scanned, a skipped …

029: Mjs20 re-anchored twice -- an unused drain and then an unused variable were compiler kills; drains zero bytes with the length kept in use, killed by the legacy-frame test

029 round 2, lane a's four: the moduledoc's cost band is the page's (2.0-2.6x, not 1.0); the citation reader sees every ExUnit skip spelling -- a tag or describetag naming skip among the attribute …

029 round 2 mutants: the reader reading one attribute line, the moduletag ignored -- zero survivors, one re-anchored past a compiler kill

029 round 2 red, lane b's findings: a header line of a legacy Content-Length block bounded like any line (today read unbounded -- 64 MiB accepted); the loop's heap about the line's bytes (today 92 …

029 round 2 green, lane b's findings: every header line of a legacy Content-Length block is read under the line bound, and past it the block and its declared body are drained (64 MiB on one header …

029 round 2: the heap pin samples the loop while it reads (what it retains after is nothing under either design, so that was no pin), and found a second forty -- the legacy Content-Length check dow…

029 round 3, lane a's low: the nesting row's heap figures are MiB as labelled -- 6,508 KB and 10,907 KB are 6.4 and 10.7 MiB, not the decimal 6.5 and 10.9 that stood under the label; the ratio 1.7x…

029 round 4, lane b: the nesting row's heap sentence compared a sampled peak with a resting heap and called the quotient 1.7x -- like for like, the bounded decode's sampled peak on the 60,000-key o…

029 round 5, lane a: the after-return heap clause dropped -- the two lanes read it two ways, because what a process holds after a decode returns is when it last collected, not a property of the dec…