Content Quality: Analysis, 1,506 words (range 800-2000), 10 sources, thesis beyond a tabulation: four operators each describe rising submissions but take four different steps, the sources word the cause differently (three of four name AI by the Herald's count, Google's quoted wording is 'automated'), and none of the cited sources gives the share produced with AI. Facts, outlet characterisations and the Herald's readings are labelled separately throughout (operator's own post vs outlet vs 'the Herald's reading'); a closing section lists what the sources do not establish. FIT: the writer's concession is fair. The assigned theme was 'AI for software development'; this piece is about bug bounty programs (vulnerability disclosure for software vendors and open-source projects) and AI's role in the report flood. It is on topic for the Herald (Cybersecurity, open-source maintenance, AI tooling side effects) and is honest that AI's effect is attributed by operators or outlets, not measured. It is only partly about AI's effect (Apple, Google and Microsoft attributions are second-hand or hedged; GitHub's is first-hand), which the article itself says. No Archive.org captures are cited (the 'earliest Archive capture' remark in the brief does not appear in the article), so the archive-citation rule does not apply.
Source Verification: All 10 snapshots read from disk (sources/2026-11/four-bug-bounty-.../source-0..9.html.gz); sha256 of every decompressed snapshot matches manifest.json; all HTTP 200, archive_fallback null, suspicious_patterns null for all ten (no prompt-injection findings; nothing to judge). Every quoted string checked verbatim on its page. source-0 GitHub blog May 15 2026 (Jarom Brown): 'a sharp increase in submissions that don't demonstrate real security impact', heading 'We welcome AI in security research', 'The tools don't matter. The quality of the work does.', validate 'scanners, static analysis, AI assistants' - confirmed. source-1 GitHub blog July 22 (Catherine Cassell): public table $250/$2,000/$5,000/$10,000, static amounts instead of ranges, VIP $1,000/$7,500/$20,000/$30,000+, 'To reduce the volume of low-effort and AI-generated reports, we're implementing a HackerOne signal requirement', 'HackerOne's platform gives researchers who don't meet the threshold up to four initial submissions', reports on or after July 27, 2026 - confirmed (so 'tiered pay' = public plus invite-only VIP tier; the article avoids the old-versus-new comparison on which the 2026-07/23 article carries a correction). source-2 The Register July 23 (Carly Page): 'Microsoft-owned code forge', motive restates GitHub's statement - confirmed; it also gives old ranges, which the article does not use. source-3 MacRumors Aug 4: 'according to The Financial Times', Apple 'limited the number of vulnerabilities security researchers can submit', 'cap on the number of open submissions a researcher can have', researchers 'can request an increase', 'amateur bug hunters using AI', Bynario ChatGPT, more than 50 macOS bugs in three weeks, Apple 'too has turned to AI for parsing submissions' - confirmed; no mention of a cooldown. source-4 The Decoder Aug 2 (published_time 2026-08-02T12:42:49Z): 'capping the number of bug reports ... and enforcing a 30-day cooldown period', cites the Financial Times, Bynario used ChatGPT, 'Researchers can request a higher quota' - confirmed; the two outlets restate one FT report (FT page 403 to me too, so also unread) = one underlying account; 'capped' matches the outlets' own wording (cap on open submissions per researcher); no limit number or effective date in either. source-5 Microsoft MSRC post August 3: 'more than $20 million to 562 security researchers', '$17 million to 344', 'a notable increase in submission volume during the second half of the year, reflecting both strong engagement ... and the growing use of AI to support security research', Zero Day Quest 'nearly 700 vulnerability reports', 'more than 300 additional reports' from the wider scope - confirmed; the post does not define 'the year'. source-6 The Register Aug 4 (Connor Jones): program period July 1, 2025 to June 30, 2026, 'In Scope By Default' (December 2025), 'You could argue that the numbers do not represent a fair fight', AI attribution attributed to Microsoft - confirmed; the January-June 2026 reading is the Herald's derived reading, labelled as such, and is arithmetically consistent. source-7 Tom's Hardware Oct 3: 'over an influx of invalid AI-driven reports', 'reportedly' overwhelmed by 'thousands' of poorly written reports, 'official X post on October 1', Intel 'did not officially confirm AI-generated reports as the reason for the move, but experts suspect this is the case' - confirmed and correctly attributed to Tom's Hardware only (Intel not counted among the four; the Phoronix and HackerOne sources were dropped, none cited). source-8 The Hacker News Oct 6: effective Oct 1, supply-chain reports and earlier reports unaffected, Q1 2027 update, Go/Angular/Protocol Buffers, quote 'a significant rise in automated submissions, the vast majority of which are not valid' attributed to a post on X, 'The post gave no figures. It did not say whether the submissions were produced with AI tools.', Angular policy 'names no other channel' - confirmed. source-9 InfoWorld Mar 20 2026: lead 'Google will no longer accept AI-generated submissions', 'increasingly concerned about the low quality of some AI-generated bug submissions', quote 'we will now require higher-quality proof (like OSS-Fuzz reproduction or a merged patch) for certain tiers' - confirmed; the lead/quote mismatch the article points out is real. EXTRA ROUTES (Google primary text): Google VRP's X post 2105689195180179605 (syndication endpoint, created 2026-10-01T16:00:00Z, response sha256 3b1512d0e8717503...) reads in full 'We are temporarily no longer accepting OSS VRP product vulnerability submissions. This does not impact OSS VRP supply chain reports, or any outstanding reports. As an alternative, we encourage you to find impact across our other VRP programs' - no reason and no 'automated' wording, as the writer reported. The rules notice in Google's public GitHub copy (google/bughunters commit f8bf23ad, Sept 30; diff sha256 25b4d02b...) reads 'As of October 1, 2026, we are no longer accepting product vulnerabilities submitted to the OSS VRP ... commit to giving an update in Q1 2027 ...' - also no reason and no quote. The quote THN attributes to Google therefore could not be traced to either primary text (the article hedges 'the quote may come from the notice or a follow-up post'; the notice text is now read and lacks it) - recorded as a clarification. The live rules page is JavaScript-only (Archive captures are 2 KB shells). Absence/quantifier sentences tested against all ten snapshots: no cited source gives an AI share (the only 'vast majority' is Google's reported claim about automated submissions being invalid, not an AI share) - holds; 'GitHub, Apple and Microsoft name AI, Google 'automated'' is the Herald's labelled count and matches each source; 'Apple's exact limit ... not given' holds. Dates recomputed: GitHub July 22 to Microsoft August 3 = 12 days (within two weeks); March 20 is InfoWorld's report date, not Google's own change date (clarification). Minor attribution: article says 'GitHub gives newcomers up to four initial submissions' in Implications while GitHub's post says HackerOne's platform gives them (clarification). INDEPENDENCE (Analysis): independent of every subject by outlet: The Register (one outlet, two cases), MacRumors and The Decoder (one underlying account, the FT, hence 1 for Apple), Tom's Hardware, The Hacker News, InfoWorld = 6 outlets, 5 underlying accounts, none a program or vendor page (the 3 vendor/program pages - 2 GitHub, 1 Microsoft - are not counted). Per case: GitHub 1 independent outlet (The Register, described as restating GitHub's motive), Apple 1 underlying account via 2 outlets (the article says so), Microsoft 1 outlet (The Register again; described as restating Microsoft's AI attribution, with its scope analysis labelled as its own), Google 3 outlets (Tom's Hardware, The Hacker News, InfoWorld, whose March report is a different event). Requirement of at least 2 independent met overall (6 outlets / 5 accounts) and the article states per case where a case rests on one outlet or account. Allowlist: all 8 hosts (github.blog, theregister.com, macrumors.com, the-decoder.com, microsoft.com, tomshardware.com, thehackernews.com, infoworld.com) already listed; no additions. Anthropic disclosure: not required - the article does not mention Anthropic, Claude or Bun (grep count 0); the MacRumors and Decoder pages mention Anthropic's models in passing but the article does not carry that material.