diff --git a/fuzz_test.go b/fuzz_test.go
new file mode 100644
index 0000000..cdd8a47
--- /dev/null
+++ b/fuzz_test.go
@@ -0,0 +1,173 @@
+package reader_test
+
+import (
+	"os"
+	"path/filepath"
+	"testing"
+	"time"
+
+	"github.com/go-pdfkit/reader"
+)
+
+// seedDir is where an adversarial corpus can be put: point PDF_SEEDS at
+// mozilla's pdf.js test suite, every file of which is there because it broke a
+// reader once, and the fuzzer starts from a far better population than
+// anything a generator would invent. Without it the committed seeds and the
+// crashers under testdata still run — the corpus makes the search better, it
+// is not what makes the test valid.
+var seedDir = os.Getenv("PDF_SEEDS")
+
+// addSeeds feeds the fuzzer files from the corpus, capped: a seed the fuzzer
+// cannot mutate quickly is a seed it will not explore.
+func addSeeds(f *testing.F, max int, cap int) {
+	f.Add([]byte("%PDF-1.7\n1 0 obj<</Type/Catalog>>endobj\ntrailer<</Root 1 0 R>>"))
+	f.Add([]byte("%PDF-1.4\nstartxref\n0\n%%EOF"))
+	if seedDir == "" {
+		return
+	}
+	ents, err := os.ReadDir(seedDir)
+	if err != nil {
+		return
+	}
+	type ent struct {
+		path string
+		size int64
+	}
+	var list []ent
+	for _, e := range ents {
+		if e.IsDir() || filepath.Ext(e.Name()) != ".pdf" {
+			continue
+		}
+		info, err := e.Info()
+		if err != nil {
+			continue
+		}
+		if info.Size() > int64(cap) {
+			continue
+		}
+		list = append(list, ent{filepath.Join(seedDir, e.Name()), info.Size()})
+	}
+	n := 0
+	for _, e := range list {
+		b, err := os.ReadFile(e.path)
+		if err != nil {
+			continue
+		}
+		f.Add(b)
+		n++
+		if n >= max {
+			break
+		}
+	}
+}
+
+// budget fails the run if one input takes longer than a reader has any right
+// to take. The fuzzer's own -timeout kills the whole process after ten
+// minutes, which tells you nothing about which input was at fault; this names
+// it. What is being hunted is a small file that costs a large amount of time,
+// so the check has to be per input.
+func budget(t *testing.T, limit time.Duration, size int, what string, f func()) {
+	t.Helper()
+	start := time.Now()
+	f()
+	if d := time.Since(start); d > limit {
+		t.Fatalf("%s: %d bytes took %s, over the %s budget", what, size, d, limit)
+	}
+}
+
+// FuzzOpenWithinABudget is the public-API twin of the in-package FuzzOpen: it
+// asserts not only that no byte string makes the reader panic, but that none
+// takes longer than a couple of seconds. Time is a return code in disguise —
+// the 219-byte file that took 21 seconds to open panicked at nothing and
+// allocated nothing, so only a clock could have caught it.
+//
+// Two fuzz targets of the same name in the two test packages cannot both be
+// selected, and `go test -fuzz FuzzOpen` refuses the ambiguity outright, so
+// neither could be run until this one was renamed.
+func FuzzOpenWithinABudget(f *testing.F) {
+	addSeeds(f, 400, 40*1024)
+	f.Fuzz(func(t *testing.T, b []byte) {
+		budget(t, 2*time.Second, len(b), "Open", func() {
+			d, err := reader.Open(b)
+			if err != nil {
+				return
+			}
+			n := d.PageCount()
+			if n > 4 {
+				n = 4
+			}
+			for i := 1; i <= n; i++ {
+				if _, err := d.Page(i); err != nil {
+					continue
+				}
+				if _, err := d.PageContent(i); err != nil {
+					continue
+				}
+				_, _ = d.PageOperations(i)
+			}
+		})
+	})
+}
+
+func FuzzParseObject(f *testing.F) {
+	for _, s := range []string{
+		"<< /A 1 /B [1 2 3] >>", "(hello \\( world)", "<AABBCC>", "12 0 R",
+		"[[[[[[[[[[]]]]]]]]]]", "<</A<</B<</C<</D 1>>>>>>>>", "3.14", "true",
+		"/Name#20with#20spaces", "null", "-.0000000000000000001",
+	} {
+		f.Add([]byte(s))
+	}
+	f.Fuzz(func(t *testing.T, b []byte) {
+		budget(t, time.Second, len(b), "ParseObject", func() {
+			o, _, err := reader.ParseObject(b)
+			if err == nil {
+				_ = reader.FormatObject(o)
+			}
+		})
+	})
+}
+
+func FuzzOperations(f *testing.F) {
+	for _, s := range []string{
+		"BT /F1 12 Tf (hi) Tj ET", "q 1 0 0 1 0 0 cm Q",
+		"BI /W 2 /H 2 /BPC 8 /CS /G ID \x00\x01\x02\x03 EI",
+		"[ (a) -250 (b) ] TJ", "1 2 3 4 5 6 c",
+	} {
+		f.Add([]byte(s))
+	}
+	f.Fuzz(func(t *testing.T, b []byte) {
+		budget(t, time.Second, len(b), "Operations", func() {
+			_, _ = reader.Operations(b)
+		})
+	})
+}
+
+// FuzzDecode drives the filter chain directly. The first byte picks the
+// filter, so that one corpus explores all of them rather than one each.
+func FuzzDecode(f *testing.F) {
+	names := []reader.Name{"FlateDecode", "LZWDecode", "ASCII85Decode", "ASCIIHexDecode", "RunLengthDecode"}
+	for i := range names {
+		f.Add(byte(i), byte(0), []byte("x\x9c\x03\x00\x00\x00\x00\x01"))
+		f.Add(byte(i), byte(2), []byte("~>"))
+	}
+	f.Fuzz(func(t *testing.T, which byte, colors byte, data []byte) {
+		name := names[int(which)%len(names)]
+		d := reader.Dict{"Filter": name}
+		if colors&1 != 0 {
+			// A predictor is where a decoder starts trusting numbers the file
+			// chose: rows, columns, colours and bits per component all come
+			// from the document, and their product is a buffer size.
+			d["DecodeParms"] = reader.Dict{
+				"Predictor":        reader.Integer(2 + int(colors)%14),
+				"Colors":           reader.Integer(1 + int(colors>>1)%8),
+				"Columns":          reader.Integer(1 + int(colors>>2)%64),
+				"BitsPerComponent": reader.Integer([]int{1, 2, 4, 8, 16}[int(colors>>3)%5]),
+			}
+		}
+		budget(t, 2*time.Second, len(data), "Decode", func() {
+			_, _, _ = reader.Decode(d, data, func(reader.Ref) (reader.Object, error) {
+				return nil, nil
+			})
+		})
+	})
+}